developer-tools Static interface comparison
Docmancer
0.9.15 0.9.16
3 schema changes require review.No new authority detected.
View the upgrade review as Markdown
A shareable summary of this comparison for an issue, pull request or ticket. Select the text and copy it.
INTERFACE DIFF
What changed, tool by tool
- 0Added tools
- 0Removed tools
- 19Modified tools
- 3Input-schema changes
- 0Output-schema changes
- 0Declared authority changes
19 tools with recorded changes
ask_memoryModifiedWhat changed
- ask_memory — Schema change: input property agent changed: description "Target agent ident… → "Target agent ident….
- ask_memory: description changed.
Why it could matter
- Calls or consumers built for the earlier schema may need to change.
- Models choose and call tools partly from their descriptions.
What to verify
- Validate representative existing tool calls against the new argument schema.
- Spot-check that agents still select this tool as intended.
- Declared effect
- read → read
- Description
- PreviousREAD-ONLY: recall one bounded bundle of everything local memory knows about a task, combining mandatory policy, curated memory, and supporting harvested agent evidence in priority order. This is the default recall tool and the right first call when you do not already know which store holds the answer; search_memory and search_evidence each cover only one store. Local and free by default. Returns an empty bundle rather than an error when nothing relevant exists, so an empty result means the memory is genuinely silent. Parameters: task is the natural-language task or question to recall for; answer defaults to false and returns the raw bundle, and setting it true spends a configured provider call, which may leave the machine, to produce a grounded cited answer; mode sets that answer's verbosity to concise, normal, or thorough and is ignored when answer is false; token_budget bounds the bundle size in approximate tokens (default 4000); limit caps how many supporting evidence items are considered (default 12); include_history adds superseded evidence; agent is the requesting agent identifier recorded for attribution (default mcp-client); project_path project root whose memory tree the operation applies to. Omit it to use the machine-wide tree, and always omit it when this server was started pinned to a project. Returns an object with the mandatory policy, curated memory, and evidence sections, each carrying source citations, plus the generated answer when answer is true. Example: task='how do we deploy this project', token_budget=2000.
- CurrentREAD-ONLY: recall one bounded bundle of everything local memory knows about a task, combining mandatory policy, curated memory, and supporting harvested agent evidence in priority order. This is the default recall tool and the right first call when you do not already know which store holds the answer; search_memory and search_evidence each cover only one store. Local and free by default. Returns an empty bundle rather than an error when nothing relevant exists, so an empty result means the memory is genuinely silent. Parameters: task is the natural-language task or question to recall for; answer defaults to false and returns the raw bundle, and setting it true spends a configured provider call, which may leave the machine, to produce a grounded cited answer; mode sets that answer's verbosity to concise, normal, or thorough and is ignored when answer is false; token_budget bounds the bundle size in approximate tokens (default 4000); limit caps how many supporting evidence items are considered (default 12); include_history adds superseded evidence; agent is the requesting agent identifier recorded for attribution (default mcp-client); project_path project root whose memory tree the operation applies to. Omit it to use the machine-wide tree, and always omit it when this server was started pinned to a project. Returns an object with the mandatory policy, curated memory, and evidence sections, each carrying source citations, plus the generated answer when answer is true.
Input schema
Property Previous (0.9.15) Current (0.9.16) Change agentstring, default="mcp-client"optional string, default="mcp-client"optional - Description changed
8 unchanged entries
(schema root)answerinclude_historylimitmodeproject_pathtasktoken_budget
No output schema is recorded for this tool in either release.
Link to this toolExact JSON from the published snapshots
Loads the exact recorded definitions of this tool from both releases. The full snapshots are linked under Evidence and provenance.
canonical_memoryModifiedWhat changed
- canonical_memory: description changed.
Why it could matter
- Models choose and call tools partly from their descriptions.
What to verify
- Spot-check that agents still select this tool as intended.
- Declared effect
- read → read
- Description
- PreviousREAD-ONLY: read the machine-wide canonical memory, which is what docmancer has reconciled about this user across every agent and project. Local disk read, machine-wide, so it takes no project_path. Call it before asking the user something they may already have told another agent. Not a search tool: it returns whole prepared sections rather than matches, so use search_memory or ask_memory to look something up. To change what it returns, use pin_memory rather than edit_memory. Parameters: section selects one of about (who the user is), preferences (how they want to work), working-principles (cross-project rules), active-projects (what they are working on), or canonical-memory (a description of the store itself); omit section entirely for a status summary of every section. Returns, with a section, that section split into its pinned zone (durable, survives reconciliation) and its generated zone (rewritten automatically), plus content_hash and revision id. Without a section, returns per-section presence and pinned-line counts. Example: section='preferences'.
- CurrentREAD-ONLY: read the machine-wide canonical memory, which is what docmancer has reconciled about this user across every agent and project. Local disk read, machine-wide, so it takes no project_path. Call it before asking the user something they may already have told another agent. Not a search tool: it returns whole prepared sections rather than matches, so use search_memory or ask_memory to look something up. To change what it returns, use pin_memory rather than edit_memory. Parameters: section selects one of about (who the user is), preferences (how they want to work), working-principles (cross-project rules), active-projects (what they are working on), or canonical-memory (a description of the store itself); omit section entirely for a status summary of every section. Returns, with a section, that section split into its pinned zone (durable, survives reconciliation) and its generated zone (rewritten automatically), plus content_hash and revision id. Without a section, returns per-section presence and pinned-line counts.
Input schema
Input schema unchanged.
2 unchanged entries
(schema root)section
No output schema is recorded for this tool in either release.
Link to this toolExact JSON from the published snapshots
Loads the exact recorded definitions of this tool from both releases. The full snapshots are linked under Evidence and provenance.
common_memoryModifiedWhat changed
- common_memory: description changed.
Why it could matter
- Models choose and call tools partly from their descriptions.
What to verify
- Spot-check that agents still select this tool as intended.
- Declared effect
- read → read
- Description
- PreviousREAD-ONLY: list the memories that recur across two or more independent agent harnesses, for example something both Claude Code and Cursor recorded separately. Local disk read over already-harvested evidence; docmancer's own generated integration copies are excluded so they cannot manufacture agreement. Recurrence is evidence of salience, not proof of correctness, so treat the result as a signal worth checking rather than settled truth. Takes no query: it returns the whole recurring set. Use search_evidence to look something specific up instead. Parameters: project_path project root whose memory tree the operation applies to. Omit it to use the machine-wide tree, and always omit it when this server was started pinned to a project. Returns a list of objects describing each recurring memory and the harnesses it was seen in. Returns [] when nothing recurs. Example: call with no arguments.
- CurrentREAD-ONLY: list the memories that recur across two or more independent agent harnesses. This is a local disk read over already-harvested evidence; docmancer's own generated integration copies are excluded so they cannot manufacture agreement. Recurrence is evidence of salience, not proof of correctness, so treat the result as a signal worth checking rather than settled truth. Takes no query: it returns the whole recurring set. Use search_evidence to look something specific up instead. Parameters: project_path project root whose memory tree the operation applies to. Omit it to use the machine-wide tree, and always omit it when this server was started pinned to a project. Returns a list of objects describing each recurring memory and the harnesses it was seen in. Returns [] when nothing recurs.
Input schema
Input schema unchanged.
2 unchanged entries
(schema root)project_path
Output schema
Output schema unchanged.
2 unchanged entries
(schema root)result
Link to this toolExact JSON from the published snapshots
Loads the exact recorded definitions of this tool from both releases. The full snapshots are linked under Evidence and provenance.
context_deliveryModifiedWhat changed
- context_delivery: description changed.
Why it could matter
- Models choose and call tools partly from their descriptions.
What to verify
- Spot-check that agents still select this tool as intended.
- Declared effect
- read → read
- Description
- PreviousREAD-ONLY: show, per agent, whether context is actually reaching it, listing each supported agent's integration mode, hook installation status, and the revision and hash of the last context bundle it was observed to receive. Local disk read with no side effects. This is the delivery question, so use it to diagnose why one agent seems to be missing context that another has. It is one of three context tools: context_status answers what the current context revision contains and how fresh it is, and context_projection renders the actual text one agent would receive. Read-only across all three; installing hooks and refreshing context are deliberately human-only CLI or local-web operations. Parameters: project_path project root whose memory tree the operation applies to. Omit it to use the machine-wide tree, and always omit it when this server was started pinned to a project. Returns a list with one row per agent giving its integration mode, hook status, and last delivered revision and hash. A stale or absent revision on one row is the signal to look for. Example: call with no arguments.
- CurrentREAD-ONLY: show, per agent, whether context is actually reaching it, listing each supported agent's integration mode, hook installation status, and the revision and hash of the last context bundle it was observed to receive. Local disk read with no side effects. This is the delivery question, so use it to diagnose why one agent seems to be missing context that another has. It is one of three context tools: context_status answers what the current context revision contains and how fresh it is, and context_projection renders the actual text one agent would receive. Read-only across all three; installing hooks and refreshing context are deliberately human-only CLI or local-web operations. Parameters: project_path project root whose memory tree the operation applies to. Omit it to use the machine-wide tree, and always omit it when this server was started pinned to a project. Returns a list with one row per agent giving its integration mode, hook status, and last delivered revision and hash. A stale or absent revision on one row is the signal to look for.
Input schema
Input schema unchanged.
2 unchanged entries
(schema root)project_path
Output schema
Output schema unchanged.
2 unchanged entries
(schema root)result
Link to this toolExact JSON from the published snapshots
Loads the exact recorded definitions of this tool from both releases. The full snapshots are linked under Evidence and provenance.
context_projectionModifiedWhat changed
- context_projection — Schema change: input property agent changed: description "Target agent ident… → "Target agent ident….
- context_projection: description changed.
Why it could matter
- Calls or consumers built for the earlier schema may need to change.
- Models choose and call tools partly from their descriptions.
What to verify
- Validate representative existing tool calls against the new argument schema.
- Spot-check that agents still select this tool as intended.
- Declared effect
- read → read
- Description
- PreviousREAD-ONLY: render the actual bounded context text one named agent would receive, linked to the current revision. Local disk read that renders without writing, refreshing, or delivering anything, so it is safe to preview repeatedly. This is the show-me-the-content question, so use it to check what an agent will actually see before installing or debugging an integration. It is one of three context tools: context_status reports the revision and its freshness, and context_delivery reports whether each agent is receiving it. Refresh, rollback, adopt, and retire are human-only CLI or local-web operations. Parameters: agent is the required target agent identifier, such as claude-code or cursor, and the projection is shaped for that agent's integration; token_budget bounds the rendered size in approximate tokens (default 2000), so raising it returns more content and lowering it truncates by priority; project_path project root whose memory tree the operation applies to. Omit it to use the machine-wide tree, and always omit it when this server was started pinned to a project. Returns an object with the rendered projection text and the revision it was built from. Example: agent='claude-code', token_budget=4000.
- CurrentREAD-ONLY: render the actual bounded context text one named agent would receive, linked to the current revision. Local disk read that renders without writing, refreshing, or delivering anything, so it is safe to preview repeatedly. This is the show-me-the-content question, so use it to check what an agent will actually see before installing or debugging an integration. It is one of three context tools: context_status reports the revision and its freshness, and context_delivery reports whether each agent is receiving it. Refresh, rollback, adopt, and retire are human-only CLI or local-web operations. Parameters: agent is the required target agent identifier, and the projection is shaped for that agent's integration; token_budget bounds the rendered size in approximate tokens (default 2000), so raising it returns more content and lowering it truncates by priority; project_path project root whose memory tree the operation applies to. Omit it to use the machine-wide tree, and always omit it when this server was started pinned to a project. Returns an object with the rendered projection text and the revision it was built from.
Input schema
Property Previous (0.9.15) Current (0.9.16) Change agentstringrequired stringrequired - Description changed
3 unchanged entries
(schema root)project_pathtoken_budget
No output schema is recorded for this tool in either release.
Link to this toolExact JSON from the published snapshots
Loads the exact recorded definitions of this tool from both releases. The full snapshots are linked under Evidence and provenance.
context_statusModifiedWhat changed
- context_status: description changed.
Why it could matter
- Models choose and call tools partly from their descriptions.
What to verify
- Spot-check that agents still select this tool as intended.
- Declared effect
- read → read
- Description
- PreviousREAD-ONLY: inspect the consolidated Context itself, reporting its current revision id, how fresh it is, which sources are excluded, and its cluster metadata. Local disk read with no side effects. This is the what-and-how-fresh question, so use it to decide whether the context an agent is working from is stale or is missing something on purpose. It is one of three context tools: context_delivery answers whether each agent is actually receiving it, and context_projection renders the actual text one agent would receive. Refresh, rollback, adopt, and retire are deliberately human-only CLI or local-web operations, so no MCP tool can change what this reports. Parameters: project_path project root whose memory tree the operation applies to. Omit it to use the machine-wide tree, and always omit it when this server was started pinned to a project. Returns an object with the current revision, its scope and freshness, the active exclusions, and cluster metadata. Example: call with no arguments.
- CurrentREAD-ONLY: inspect the consolidated Context itself, reporting its current revision id, how fresh it is, which sources are excluded, and its cluster metadata. Local disk read with no side effects. This is the what-and-how-fresh question, so use it to decide whether the context an agent is working from is stale or is missing something on purpose. It is one of three context tools: context_delivery answers whether each agent is actually receiving it, and context_projection renders the actual text one agent would receive. Refresh, rollback, adopt, and retire are deliberately human-only CLI or local-web operations, so no MCP tool can change what this reports. Parameters: project_path project root whose memory tree the operation applies to. Omit it to use the machine-wide tree, and always omit it when this server was started pinned to a project. Returns an object with the current revision, its scope and freshness, the active exclusions, and cluster metadata.
Input schema
Input schema unchanged.
2 unchanged entries
(schema root)project_path
No output schema is recorded for this tool in either release.
Link to this toolExact JSON from the published snapshots
Loads the exact recorded definitions of this tool from both releases. The full snapshots are linked under Evidence and provenance.
decision_timelineModifiedWhat changed
- decision_timeline: description changed.
Why it could matter
- Models choose and call tools partly from their descriptions.
What to verify
- Spot-check that agents still select this tool as intended.
- Declared effect
- read → read
- Description
- PreviousREAD-ONLY: show the append-only history of how memory changed over time, with a human-readable diff per entry. Local disk read over a log that is only ever appended to, so nothing here can be rewritten and the record is trustworthy. Use it to answer when and how a memory changed, or to recover text that an edit replaced. Newest entries come first. This is history, not content: use search_memory or read_memory for what a file says now. Parameters: file_id restricts the timeline to one memory file's own history, which is the usual way to trace a single decision; operation restricts it to one kind of change, one of create, edit, move, duplicate, trash, restore, pin, or reconcile; limit caps how many entries come back (default 100); project_path project root whose memory tree the operation applies to. Omit it to use the machine-wide tree, and always omit it when this server was started pinned to a project. Omit every filter for the whole recent history. Returns a list of entries with the operation, timestamp, affected file, revision ids, and a readable diff. Example: file_id='01J8XY...', operation='edit', limit=20.
- CurrentREAD-ONLY: show the append-only history of how memory changed over time, with a human-readable diff per entry. Local disk read over a log that is only ever appended to, so nothing here can be rewritten and the record is trustworthy. Use it to answer when and how a memory changed, or to recover text that an edit replaced. Newest entries come first. This is history, not content: use search_memory or read_memory for what a file says now. Parameters: file_id restricts the timeline to one memory file's own history, which is the usual way to trace a single decision; operation restricts it to one kind of change, one of create, edit, move, duplicate, trash, restore, pin, or reconcile; limit caps how many entries come back (default 100); project_path project root whose memory tree the operation applies to. Omit it to use the machine-wide tree, and always omit it when this server was started pinned to a project. Omit every filter for the whole recent history. Returns a list of entries with the operation, timestamp, affected file, revision ids, and a readable diff.
Input schema
Input schema unchanged.
5 unchanged entries
(schema root)file_idlimitoperationproject_path
Output schema
Output schema unchanged.
2 unchanged entries
(schema root)result
Link to this toolExact JSON from the published snapshots
Loads the exact recorded definitions of this tool from both releases. The full snapshots are linked under Evidence and provenance.
duplicate_memoryModifiedWhat changed
- duplicate_memory: description changed.
Why it could matter
- Models choose and call tools partly from their descriptions.
What to verify
- Spot-check that agents still select this tool as intended.
- Declared effect
- write → write
- Description
- PreviousMUTATING: copy one curated memory file to a new path under a new stable identity. Writes to local disk and appends to the change timeline. The original is left untouched, and the copy gets its own docmancer://memory address and its own history, so the two diverge from here and editing one does not affect the other. Use it to fork an existing memory into a variant. Use move_memory instead when the original should not survive. Not idempotent: calling it twice with the same new_relative_path fails on the second call rather than creating a second copy. Parameters: address identifies the file to copy; new_relative_path is the copy's path in the tree, including the .md suffix, and must not already exist; expected_hash is the source file's current content_hash from read_memory, which guards against copying a revision you have not seen; project_path project root whose memory tree the operation applies to. Omit it to use the machine-wide tree, and always omit it when this server was started pinned to a project. Returns an object with the new copy's address, path, content_hash, and revision id. Example: address='docmancer://memory/01J8XY...', new_relative_path='decisions/hosting-staging.md', expected_hash='9f2c...'.
- CurrentMUTATING: copy one curated memory file to a new path under a new stable identity. Writes to local disk and appends to the change timeline. The original is left untouched, and the copy gets its own docmancer://memory address and its own history, so the two diverge from here and editing one does not affect the other. Use it to fork an existing memory into a variant. Use move_memory instead when the original should not survive. Not idempotent: calling it twice with the same new_relative_path fails on the second call rather than creating a second copy. Parameters: address identifies the file to copy; new_relative_path is the copy's path in the tree, including the .md suffix, and must not already exist; expected_hash is the source file's current content_hash from read_memory, which guards against copying a revision you have not seen; project_path project root whose memory tree the operation applies to. Omit it to use the machine-wide tree, and always omit it when this server was started pinned to a project. Returns an object with the new copy's address, path, content_hash, and revision id.
Input schema
Input schema unchanged.
5 unchanged entries
(schema root)addressexpected_hashnew_relative_pathproject_path
No output schema is recorded for this tool in either release.
Link to this toolExact JSON from the published snapshots
Loads the exact recorded definitions of this tool from both releases. The full snapshots are linked under Evidence and provenance.
edit_memoryModifiedWhat changed
- edit_memory: description changed.
Why it could matter
- Models choose and call tools partly from their descriptions.
What to verify
- Spot-check that agents still select this tool as intended.
- Declared effect
- write → write
- Description
- PreviousMUTATING: replace the body of one existing curated memory file while preserving its frontmatter, address, and history. Writes to local disk and appends to the change timeline. The write is guarded: a stale expected_hash fails without changing anything and returns a structured error naming the re-read-and-retry next action, so concurrent edits cannot be lost. Read the file first to obtain the hash. Not for canonical memory about the user, which reconciliation regenerates: use pin_memory there. Not for creating a new file: use write_memory. Parameters: address identifies the file to edit; text is the complete replacement Markdown body, which overwrites rather than appends; expected_hash is the content_hash from the read that produced text; project_path project root whose memory tree the operation applies to. Omit it to use the machine-wide tree, and always omit it when this server was started pinned to a project. Returns an object with the address and the new content_hash and revision id. Example: address='docmancer://memory/01J8XY...', text='# Deployment\n\nWe use Fly.io.', expected_hash='9f2c...'.
- CurrentMUTATING: replace the body of one existing curated memory file while preserving its frontmatter, address, and history. Writes to local disk and appends to the change timeline. The write is guarded: a stale expected_hash fails without changing anything and returns a structured error naming the re-read-and-retry next action, so concurrent edits cannot be lost. Read the file first to obtain the hash. Not for canonical memory about the user, which reconciliation regenerates: use pin_memory there. Not for creating a new file: use write_memory. Parameters: address identifies the file to edit; text is the complete replacement Markdown body, which overwrites rather than appends; expected_hash is the content_hash from the read that produced text; project_path project root whose memory tree the operation applies to. Omit it to use the machine-wide tree, and always omit it when this server was started pinned to a project. Returns an object with the address and the new content_hash and revision id.
Input schema
Input schema unchanged.
5 unchanged entries
(schema root)addressexpected_hashproject_pathtext
No output schema is recorded for this tool in either release.
Link to this toolExact JSON from the published snapshots
Loads the exact recorded definitions of this tool from both releases. The full snapshots are linked under Evidence and provenance.
move_memoryModifiedWhat changed
- move_memory: description changed.
Why it could matter
- Models choose and call tools partly from their descriptions.
What to verify
- Spot-check that agents still select this tool as intended.
- Declared effect
- write → write
- Description
- PreviousMUTATING and DESTRUCTIVE at the old path: move or rename one curated memory file. Writes to local disk and appends to the change timeline. The file's stable docmancer://memory address survives the move, so existing references by address keep working, but the old relative path stops resolving and anything referring to it by path breaks. Guarded by expected_hash, so a stale hash fails without changing anything. The body is untouched: use edit_memory to change content, or duplicate_memory to copy rather than move. Parameters: address identifies the file to move; new_relative_path is its destination path in the tree, including the .md suffix, and renaming is just a move within the same directory; expected_hash is the file's current content_hash from read_memory; project_path project root whose memory tree the operation applies to. Omit it to use the machine-wide tree, and always omit it when this server was started pinned to a project. Returns an object with the unchanged address, the new path, and the new revision id. Example: address='docmancer://memory/01J8XY...', new_relative_path='decisions/hosting.md', expected_hash='9f2c...'.
- CurrentMUTATING and DESTRUCTIVE at the old path: move or rename one curated memory file. Writes to local disk and appends to the change timeline. The file's stable docmancer://memory address survives the move, so existing references by address keep working, but the old relative path stops resolving and anything referring to it by path breaks. Guarded by expected_hash, so a stale hash fails without changing anything. The body is untouched: use edit_memory to change content, or duplicate_memory to copy rather than move. Parameters: address identifies the file to move; new_relative_path is its destination path in the tree, including the .md suffix, and renaming is just a move within the same directory; expected_hash is the file's current content_hash from read_memory; project_path project root whose memory tree the operation applies to. Omit it to use the machine-wide tree, and always omit it when this server was started pinned to a project. Returns an object with the unchanged address, the new path, and the new revision id.
Input schema
Input schema unchanged.
5 unchanged entries
(schema root)addressexpected_hashnew_relative_pathproject_path
No output schema is recorded for this tool in either release.
Link to this toolExact JSON from the published snapshots
Loads the exact recorded definitions of this tool from both releases. The full snapshots are linked under Evidence and provenance.
pin_memoryModifiedWhat changed
- pin_memory: description changed.
Why it could matter
- Models choose and call tools partly from their descriptions.
What to verify
- Spot-check that agents still select this tool as intended.
- Declared effect
- write → write
- Description
- PreviousMUTATING: pin one durable line into a canonical memory section. Writes to local disk, machine-wide, so it takes no project_path. The pinned zone is the ONLY part of a canonical section that survives automatic reconciliation, so use this, not edit_memory or write_memory, for any correction, standing preference, or fact the reconciler got wrong or left out. Idempotent: pinning the same line twice changes nothing. Use unpin_memory to reverse it, and canonical_memory to see the result. Parameters: section selects which canonical section to pin into, one of about, preferences, working-principles, or active-projects; text is the complete line to pin, which should read as a standalone statement because it is stored verbatim. Returns an object with the section, its path, and the updated pinned-line count. Example: section='preferences', text='Prefers pnpm over npm for all TypeScript work.'
- CurrentMUTATING: pin one durable line into a canonical memory section. Writes to local disk, machine-wide, so it takes no project_path. The pinned zone is the ONLY part of a canonical section that survives automatic reconciliation, so use this, not edit_memory or write_memory, for any correction, standing preference, or fact the reconciler got wrong or left out. Idempotent: pinning the same line twice changes nothing. Use unpin_memory to reverse it, and canonical_memory to see the result. Parameters: section selects which canonical section to pin into, one of about, preferences, working-principles, or active-projects; text is the complete line to pin, which should read as a standalone statement because it is stored verbatim. Returns an object with the section, its path, and the updated pinned-line count.
Input schema
Input schema unchanged.
3 unchanged entries
(schema root)sectiontext
No output schema is recorded for this tool in either release.
Link to this toolExact JSON from the published snapshots
Loads the exact recorded definitions of this tool from both releases. The full snapshots are linked under Evidence and provenance.
read_memoryModifiedWhat changed
- read_memory: description changed.
Why it could matter
- Models choose and call tools partly from their descriptions.
What to verify
- Spot-check that agents still select this tool as intended.
- Declared effect
- read → read
- Description
- PreviousREAD-ONLY: read one curated memory file in full, resolved by stable address, relative path, or exact title. Local disk read with no side effects. Ambiguous title or path matches return every candidate address rather than guessing, so a caller can retry with an exact address. This is also how you obtain the content_hash that edit_memory, move_memory, duplicate_memory, and trash_memory all require. Use search_memory first when you do not already know which file you want. Parameters: address is the docmancer://memory/<id> address, the relative path, or the exact title; project_path project root whose memory tree the operation applies to. Omit it to use the machine-wide tree, and always omit it when this server was started pinned to a project. Returns an object with the body, frontmatter, address, content_hash, and revision id, or a candidates list when the address was ambiguous. Example: address='docmancer://memory/01J8XY...'.
- CurrentREAD-ONLY: read one curated memory file in full, resolved by stable address, relative path, or exact title. Local disk read with no side effects. Ambiguous title or path matches return every candidate address rather than guessing, so a caller can retry with an exact address. This is also how you obtain the content_hash that edit_memory, move_memory, duplicate_memory, and trash_memory all require. Use search_memory first when you do not already know which file you want. Parameters: address is the docmancer://memory/<id> address, the relative path, or the exact title; project_path project root whose memory tree the operation applies to. Omit it to use the machine-wide tree, and always omit it when this server was started pinned to a project. Returns an object with the body, frontmatter, address, content_hash, and revision id, or a candidates list when the address was ambiguous.
Input schema
Input schema unchanged.
3 unchanged entries
(schema root)addressproject_path
No output schema is recorded for this tool in either release.
Link to this toolExact JSON from the published snapshots
Loads the exact recorded definitions of this tool from both releases. The full snapshots are linked under Evidence and provenance.
restore_memoryModifiedWhat changed
- restore_memory: description changed.
Why it could matter
- Models choose and call tools partly from their descriptions.
What to verify
- Spot-check that agents still select this tool as intended.
- Declared effect
- write → write
- Description
- PreviousMUTATING: undo a trash_memory call by restoring one curated memory file from its restore token. Writes to local disk and appends to the change timeline. The file returns to its original path with its address and history intact, so references by address resume working. A token is single-use and only valid for the tree it was issued against, so a spent or foreign token fails without changing anything. Restoring is only possible while the token is known: there is no way to browse trash from MCP. Parameters: restore_token is the exact token returned by the trash_memory call being undone; project_path project root whose memory tree the operation applies to. Omit it to use the machine-wide tree, and always omit it when this server was started pinned to a project. Returns an object with the restored address, path, and revision id. Example: restore_token='trash-01J8XY...'.
- CurrentMUTATING: undo a trash_memory call by restoring one curated memory file from its restore token. Writes to local disk and appends to the change timeline. The file returns to its original path with its address and history intact, so references by address resume working. A token is single-use and only valid for the tree it was issued against, so a spent or foreign token fails without changing anything. Restoring is only possible while the token is known: there is no way to browse trash from MCP. Parameters: restore_token is the exact token returned by the trash_memory call being undone; project_path project root whose memory tree the operation applies to. Omit it to use the machine-wide tree, and always omit it when this server was started pinned to a project. Returns an object with the restored address, path, and revision id.
Input schema
Input schema unchanged.
3 unchanged entries
(schema root)project_pathrestore_token
No output schema is recorded for this tool in either release.
Link to this toolExact JSON from the published snapshots
Loads the exact recorded definitions of this tool from both releases. The full snapshots are linked under Evidence and provenance.
search_docsModifiedWhat changed
- search_docs: description changed.
Why it could matter
- Models choose and call tools partly from their descriptions.
What to verify
- Spot-check that agents still select this tool as intended.
- Declared effect
- read → read
- Description
- PreviousREAD-ONLY: search the local documentation index built by `docmancer docs add`, covering library, API, and vendor documentation the user chose to ingest. Runs entirely locally; no network call, so it only ever returns docs already added on this machine and returns [] when none have been. Use it for version-specific library or vendor behaviour. Not for anything about this user or their projects: use search_memory, search_evidence, or ask_memory for that. Parameters: query is the natural-language text to match against the indexed documentation; limit caps the number of results (default 8). Returns a list of objects with the matching excerpt, its source document, and a relevance score. Example: query='fastapi dependency injection', limit=5.
- CurrentREAD-ONLY: search the local documentation index built by `docmancer docs add`, covering library, API, and vendor documentation the user chose to ingest. Runs entirely locally; no network call, so it only ever returns docs already added on this machine and returns [] when none have been. Use it for version-specific library or vendor behaviour. Not for anything about this user or their projects: use search_memory, search_evidence, or ask_memory for that. Parameters: query is the natural-language text to match against the indexed documentation; limit caps the number of results (default 8). Returns a list of objects with the matching excerpt, its source document, and a relevance score.
Input schema
Input schema unchanged.
3 unchanged entries
(schema root)limitquery
Output schema
Output schema unchanged.
3 unchanged entries
(schema root)resultresult › [items]
Link to this toolExact JSON from the published snapshots
Loads the exact recorded definitions of this tool from both releases. The full snapshots are linked under Evidence and provenance.
search_evidenceModifiedWhat changed
- search_evidence: description changed.
Why it could matter
- Models choose and call tools partly from their descriptions.
What to verify
- Spot-check that agents still select this tool as intended.
- Declared effect
- read → read
- Description
- PreviousREAD-ONLY: semantic search over the raw evidence index, meaning the memory, instruction, and rule files that other coding agents already wrote on this machine and that docmancer harvested. Runs entirely locally against the embedding index; no network call, no cost. Use it to find what an agent originally recorded, with a relevance score and the file it came from. Not for curated memory you or an agent deliberately wrote: that lives in a separate tree, so use search_memory for it, or ask_memory to get both at once. Parameters: query is the natural-language text to match; limit caps the number of results (default 8); include_history adds superseded evidence to the active evidence; expand_relations adds items linked to a direct match. Returns a list of objects with score, excerpt, source_path, scope, kind, lifecycle_state, and a docmancer://record/<id> record_uri. Returns [] when nothing matches. Example: query='why did we choose Railway', limit=5.
- CurrentREAD-ONLY: semantic search over the raw evidence index, meaning the memory, instruction, and rule files that other coding agents already wrote on this machine and that docmancer harvested. Runs entirely locally against the embedding index; no network call, no cost. Use it to find what an agent originally recorded, with a relevance score and the file it came from. Not for curated memory you or an agent deliberately wrote: that lives in a separate tree, so use search_memory for it, or ask_memory to get both at once. Parameters: query is the natural-language text to match; limit caps the number of results (default 8); include_history adds superseded evidence to the active evidence; expand_relations adds items linked to a direct match. Returns a list of objects with score, excerpt, source_path, scope, kind, lifecycle_state, and a docmancer://record/<id> record_uri. Returns [] when nothing matches.
Input schema
Input schema unchanged.
5 unchanged entries
(schema root)expand_relationsinclude_historylimitquery
Output schema
Output schema unchanged.
3 unchanged entries
(schema root)resultresult › [items]
Link to this toolExact JSON from the published snapshots
Loads the exact recorded definitions of this tool from both releases. The full snapshots are linked under Evidence and provenance.
search_memoryModifiedWhat changed
- search_memory: description changed.
Why it could matter
- Models choose and call tools partly from their descriptions.
What to verify
- Spot-check that agents still select this tool as intended.
- Declared effect
- read → read
- Description
- PreviousREAD-ONLY: lexical search over the curated memory tree, meaning the decisions, constraints, and conventions deliberately written by the user or an agent. Local disk read, no embedding call, no cost. Only active entries are searched; archived and trashed files are excluded. Returns [] rather than an error when nothing is relevant or the tree is empty, so an empty result is a real answer and not a failure. Not the same store as search_evidence, which searches raw harvested agent files instead; use ask_memory when you want both stores plus mandatory policy in one bundle. Parameters: query is the natural-language text to match; limit caps the number of results (default 8, capped at 50); project_path project root whose memory tree the operation applies to. Omit it to use the machine-wide tree, and always omit it when this server was started pinned to a project. Returns a list of objects with the docmancer://memory address, title, excerpt, authority, and source_type. Pass an address to read_memory for the full file. Example: query='prior deployment decision', limit=5.
- CurrentREAD-ONLY: lexical search over the curated memory tree, meaning the decisions, constraints, and conventions deliberately written by the user or an agent. Local disk read, no embedding call, no cost. Only active entries are searched; archived and trashed files are excluded. Returns [] rather than an error when nothing is relevant or the tree is empty, so an empty result is a real answer and not a failure. Not the same store as search_evidence, which searches raw harvested agent files instead; use ask_memory when you want both stores plus mandatory policy in one bundle. Parameters: query is the natural-language text to match; limit caps the number of results (default 8, capped at 50); project_path project root whose memory tree the operation applies to. Omit it to use the machine-wide tree, and always omit it when this server was started pinned to a project. Returns a list of objects with the docmancer://memory address, title, excerpt, authority, and source_type. Pass an address to read_memory for the full file.
Input schema
Input schema unchanged.
4 unchanged entries
(schema root)limitproject_pathquery
Output schema
Output schema unchanged.
2 unchanged entries
(schema root)result
Link to this toolExact JSON from the published snapshots
Loads the exact recorded definitions of this tool from both releases. The full snapshots are linked under Evidence and provenance.
trash_memoryModifiedWhat changed
- trash_memory: description changed.
Why it could matter
- Models choose and call tools partly from their descriptions.
What to verify
- Spot-check that agents still select this tool as intended.
- Declared effect
- write → write
- Description
- PreviousDESTRUCTIVE BUT REVERSIBLE: move one curated memory file to trash so it stops being read or recalled. Writes to local disk and appends to the change timeline. Nothing is erased: the call returns a restore_token that restore_memory consumes to bring the file back, so keep that token in your reply if the user might change their mind. Guarded by expected_hash, so a stale hash fails without trashing anything. Prefer setting status='archived' with write_memory when the file should stay readable but stop being recalled. Parameters: address identifies the file to trash; expected_hash is its current content_hash from read_memory, confirming you are discarding the revision you actually saw; project_path project root whose memory tree the operation applies to. Omit it to use the machine-wide tree, and always omit it when this server was started pinned to a project. Returns an object with the trashed address and the restore_token needed to undo it. Example: address='docmancer://memory/01J8XY...', expected_hash='9f2c...'.
- CurrentDESTRUCTIVE BUT REVERSIBLE: move one curated memory file to trash so it stops being read or recalled. Writes to local disk and appends to the change timeline. Nothing is erased: the call returns a restore_token that restore_memory consumes to bring the file back, so keep that token in your reply if the user might change their mind. Guarded by expected_hash, so a stale hash fails without trashing anything. Prefer setting status='archived' with write_memory when the file should stay readable but stop being recalled. Parameters: address identifies the file to trash; expected_hash is its current content_hash from read_memory, confirming you are discarding the revision you actually saw; project_path project root whose memory tree the operation applies to. Omit it to use the machine-wide tree, and always omit it when this server was started pinned to a project. Returns an object with the trashed address and the restore_token needed to undo it.
Input schema
Input schema unchanged.
4 unchanged entries
(schema root)addressexpected_hashproject_path
No output schema is recorded for this tool in either release.
Link to this toolExact JSON from the published snapshots
Loads the exact recorded definitions of this tool from both releases. The full snapshots are linked under Evidence and provenance.
unpin_memoryModifiedWhat changed
- unpin_memory: description changed.
Why it could matter
- Models choose and call tools partly from their descriptions.
What to verify
- Spot-check that agents still select this tool as intended.
- Declared effect
- write → write
- Description
- PreviousMUTATING and DESTRUCTIVE: permanently remove pinned lines from a canonical memory section by case-insensitive substring match, with no undo and no restore token. Writes to local disk, machine-wide, so it takes no project_path. A substring can match more lines than intended, so read the section with canonical_memory first and pass text specific enough to hit only what you mean. Fails without changing anything when nothing matches, which makes a dry run safe. Only pinned lines are removable; generated content is rewritten by reconciliation instead. Parameters: section selects which canonical section to modify, one of about, preferences, working-principles, or active-projects; text is the case-insensitive substring, not a whole line and not a pattern, identifying the pinned lines to delete. Returns an object with the section and how many lines were removed. Example: section='preferences', text='pnpm'.
- CurrentMUTATING and DESTRUCTIVE: permanently remove pinned lines from a canonical memory section by case-insensitive substring match, with no undo and no restore token. Writes to local disk, machine-wide, so it takes no project_path. A substring can match more lines than intended, so read the section with canonical_memory first and pass text specific enough to hit only what you mean. Fails without changing anything when nothing matches, which makes a dry run safe. Only pinned lines are removable; generated content is rewritten by reconciliation instead. Parameters: section selects which canonical section to modify, one of about, preferences, working-principles, or active-projects; text is the case-insensitive substring, not a whole line and not a pattern, identifying the pinned lines to delete. Returns an object with the section and how many lines were removed.
Input schema
Input schema unchanged.
3 unchanged entries
(schema root)sectiontext
No output schema is recorded for this tool in either release.
Link to this toolExact JSON from the published snapshots
Loads the exact recorded definitions of this tool from both releases. The full snapshots are linked under Evidence and provenance.
write_memoryModifiedWhat changed
- write_memory — Schema change: input property memory_type changed: description "Free-form memory c… → "Free-form memory c….
- write_memory — Schema change: input property sources changed: description "Source references,… → "Source references ….
- write_memory — Schema change: input property status changed: description "Free-form lifecycl… → "Free-form lifecycl….
- write_memory: description changed.
Why it could matter
- Calls or consumers built for the earlier schema may need to change.
- Models choose and call tools partly from their descriptions.
What to verify
- Validate representative existing tool calls against the new argument schema.
- Spot-check that agents still select this tool as intended.
- Declared effect
- write → write
- Description
- PreviousMUTATING: create or update one curated memory file in the memory tree, addressed as docmancer://memory/<id>. Writes to local disk and appends to the change timeline; never silently clobbers, because the expect guard decides whether an existing file may be replaced. Use it to record a durable decision, constraint, or convention. Not for correcting the reconciled canonical memory about the user, which is regenerated and would discard the edit: use pin_memory for that. To change only the body of an existing file, prefer edit_memory. Parameters: relative_path is the file's path in the tree, such as 'decisions/deployment.md'; text is the full Markdown body; expect is the write guard, 'absent' (default) to create only and fail if the path exists, or the current content_hash to permit a guarded overwrite; memory_type, scope, authority, status, and curation_origin are frontmatter labels described in the schema; project_id names the owning project; sources lists backing file paths or URLs; tags lists retrieval labels; project_path project root whose memory tree the operation applies to. Omit it to use the machine-wide tree, and always omit it when this server was started pinned to a project. Returns an object with the stable address, the new content_hash, and the revision id, all of which later guarded calls need. Example: relative_path='decisions/deployment.md', text='# Deployment\n\nWe use Railway.', authority='mandatory'.
- CurrentMUTATING: create or update one curated memory file in the memory tree, addressed as docmancer://memory/<id>. Writes to local disk and appends to the change timeline; never silently clobbers, because the expect guard decides whether an existing file may be replaced. Use it to record a durable decision, constraint, or convention. Not for correcting the reconciled canonical memory about the user, which is regenerated and would discard the edit: use pin_memory for that. To change only the body of an existing file, prefer edit_memory. Parameters: relative_path is the file's path in the tree and must include the Markdown suffix; text is the full Markdown body; expect is the write guard, 'absent' (default) to create only and fail if the path exists, or the current content_hash to permit a guarded overwrite; memory_type, scope, authority, status, and curation_origin are frontmatter labels described in the schema; project_id names the owning project; sources lists backing file paths or URLs; tags lists retrieval labels; project_path project root whose memory tree the operation applies to. Omit it to use the machine-wide tree, and always omit it when this server was started pinned to a project. Returns an object with the stable address, the new content_hash, and the revision id, all of which later guarded calls need.
Input schema
Property Previous (0.9.15) Current (0.9.16) Change memory_typestring, default="fact"optional string, default="fact"optional - Description changed
sourcesanyOf (2), default=nulloptional anyOf (2), default=nulloptional - Description changed
statusstring, default="active"optional string, default="active"optional - Description changed
10 unchanged entries
(schema root)authoritycuration_originexpectproject_idproject_pathrelative_pathscopetagstext
No output schema is recorded for this tool in either release.
Link to this toolExact JSON from the published snapshots
Loads the exact recorded definitions of this tool from both releases. The full snapshots are linked under Evidence and provenance.
No tools match these filters.
1 tool with no recorded change
evidence_status
Evidence and provenance
- Static findings
- REVIEW from Modelara’s automated analysis of the two declared interfaces. A policy outcome, not a security rating or a runtime test.
- Human review
- Pending human review. No reviewer decision is recorded for this comparison.
- Published evidence
- Part of Valyd’s published data generation listed in valyd-data-source.json. Published records are not rewritten.
Show provenance and exact files
- Comparison
docmancer/0.9.15~0.9.16The published comparison identifier.- Analyzer
modelara v0.1.8Policy version not recorded.Analyzed Nov 16, 2026, 11:09 PM UTC.- Distribution
pypi- Previous release
docmancer@0.9.1520 tools · published by the registry Aug 9, 2026, 6:59 AM UTC · upstream 0.9.15Artifact integritysha256-jm6CV/o3KJYNmUPCraEwO1Qwt2/m32b0OFH7oSkWyKA=Artifactdocmancer-0.9.15-py3-none-any.whlInterface fingerprintsha256:2f2fb3741a9cf83a28e95b5c3aefddf3c91e9760ba83b7dfba99f050dd1d37d3Exact interface snapshot (JSON)- Current release
docmancer@0.9.1620 tools · published by the registry Aug 11, 2026, 1:43 PM UTC · upstream 0.9.16Artifact integritysha256-3HnMFfoyiCH/a20J+7K7J53ZPFjp4Sn22cZFKuLD+aY=Artifactdocmancer-0.9.16-py3-none-any.whlInterface fingerprintsha256:216a63beb9a9e1dfcf777565b1c771ed7e1378d682dd5668a6a5334abb1b5095Exact interface snapshot (JSON)- Comparison record
- Exact comparison record (JSON)
The integrity value or digest identifies the exact artifact that was inspected; it does not show that the package is safe. The interface fingerprint identifies the declared tools and schemas Modelara recorded for that release.
MODELARA ANALYSIS OUTPUT
Interface findings
Requires review
3 findingsask_memoryREVIEWChangedShow details
input property agent changed: description "Target agent ident… → "Target agent ident…
context_projectionREVIEWChangedShow details
input property agent changed: description "Target agent ident… → "Target agent ident…
write_memoryREVIEWChangedShow details
input property memory_type changed: description "Free-form memory c… → "Free-form memory c…input property sources changed: description "Source references,… → "Source references …input property status changed: description "Free-form lifecycl… → "Free-form lifecycl…
Informational
19 findingsask_memoryINFOChangedcanonical_memoryINFOChangedcommon_memoryINFOChangedcontext_deliveryINFOChangedcontext_projectionINFOChangedcontext_statusINFOChangeddecision_timelineINFOChangedduplicate_memoryINFOChangededit_memoryINFOChangedmove_memoryINFOChangedpin_memoryINFOChangedread_memoryINFOChangedrestore_memoryINFOChangedsearch_docsINFOChangedsearch_evidenceINFOChangedsearch_memoryINFOChangedtrash_memoryINFOChangedunpin_memoryINFOChangedwrite_memoryINFOChanged
What do these results mean?
- PASS
- No policy-significant interface change detected.
- REVIEW
- Interface changes require inspection.
- FAIL
- A policy-significant change was detected.
- INVALID
- The comparison could not be evaluated reliably.