CommandSuite v0.10.0

csuite-cli

Minor Changes

• #311 20cd8c1 Thanks @andrew-jon-p7a! - Breaking. process_document is renamed team_process across the agent-facing surface, the wire, the SDK and storage. "Document" named the storage shape, not the concept: it framed the team's process as a static artifact to fetch rather than operational context to consult, and neither the permission leaf (process.manage) nor the store's own description had ever used the word. Bare process was not an option — it shadows Node's global and reads as the runtime process at every call site — and team_process rules that reading out on sight.

  What moves:

  • MCP tools. process_document_get, process_document_history and process_document_write become team_process_get, team_process_history and team_process_write. Only the new names are advertised in tools/list. The old names still dispatch for this one release, because tool names are quoted in prose that does not auto-update (team instructions, the team process text itself) and a hard rename would strand a caller reading them out of it; each such call logs a warn record naming the alias and its replacement so stale references can be found. The aliases are removed in the next minor.
  • REST. GET/PUT /process-document and GET /process-document/history become /team-process and /team-process/history. The GET /instructions field processDocument becomes teamProcess, and the instruction-block kind process_document — in blocks, in instruction events and in the context watchdog's context.block.kind attribute — becomes team_process.
  • SDK. The ProcessDocument* types and schemas become TeamProcess* (TeamProcess, TeamProcessEdit, EditTeamProcessRequest, TEAM_PROCESS_MAX, TEAM_PROCESS_PATHS, …), and Client.getProcessDocument, processDocumentHistory and writeProcessDocument become getTeamProcess, teamProcessHistory and writeTeamProcess. csuite-core now exports createSqliteTeamProcessStore, TeamProcessStore and TeamProcessError.
  • Permission. process.manage becomes team_process.manage. There is no permissions migration: raw_permissions is stored verbatim and resolved on every read, and the resolver maps the old name to the new one, so a member or template saved under process.manage keeps the capability with no row rewritten. Both names are accepted on write.
  • Database. The tables process_document and process_document_edits become team_process and team_process_edits. The store renames them in place on first open, deciding from sqlite_master rather than from a caught error: a fresh database gets the new names directly, a pre-rename database is renamed with every row — including the append-only edit history — preserved, and an already-migrated database is left untouched, so opening twice is the same as opening once.

  Upgrade brokers and runners together. A runner built before this change validates the GET /instructions packet against a block-kind enum that has no team_process in it and rejects the whole response, so a new broker leaves an older runner unable to fetch its instructions at all — the same lockstep 0.9.0 asked for.

  The instruction-version hash is computed from the same inputs as before, so upgrading does not mark any runner restart-pending.

Patch Changes

• #311 20cd8c1 Thanks @andrew-jon-p7a! - Instruction and environment restarts now start the agent cold instead of resuming the prior conversation. The successor gets the refreshed instructions and the runner's context_refresh re-brief on its open objectives, and its session_start says why (resumed: false with resumeReason: 'instructions changed', 'environment changed', or 'environment reloaded'); a fresh MCP session is always re-briefed, even inside the cooldown that folds an attach and a compaction. Resuming there was the source of a capture defect measured in production: a restarted runner re-read the whole resumed transcript as new activity — 1000+ rows spanning 15 hours in one second — saturating the uploader and dropping 40–88% of events. Two fixes close that regardless of how a session restarts: the transcript reader follows a changed transcript_path (a new file after any agent swap) while keeping its line-uuid dedup, and every transcript-derived ActivityEvent carries an optional sourceId that the broker dedups on (member_activity.source_id, a partial unique index added lazily; id-less events from older runners are stored as before). The runner no longer predicts resume from the filesystem: bare --resume on csuite claude hands the decision to Claude Code (continue, which starts fresh when nothing exists), and on csuite codex starts a new thread loudly, since codex resumes by id only; explicit --resume <id> stays strict on both.

csuite-server

Minor Changes

• #311 20cd8c1 Thanks @andrew-jon-p7a! - Breaking. process_document is renamed team_process across the agent-facing surface, the wire, the SDK and storage. "Document" named the storage shape, not the concept: it framed the team's process as a static artifact to fetch rather than operational context to consult, and neither the permission leaf (process.manage) nor the store's own description had ever used the word. Bare process was not an option — it shadows Node's global and reads as the runtime process at every call site — and team_process rules that reading out on sight.

  What moves:

  • MCP tools. process_document_get, process_document_history and process_document_write become team_process_get, team_process_history and team_process_write. Only the new names are advertised in tools/list. The old names still dispatch for this one release, because tool names are quoted in prose that does not auto-update (team instructions, the team process text itself) and a hard rename would strand a caller reading them out of it; each such call logs a warn record naming the alias and its replacement so stale references can be found. The aliases are removed in the next minor.
  • REST. GET/PUT /process-document and GET /process-document/history become /team-process and /team-process/history. The GET /instructions field processDocument becomes teamProcess, and the instruction-block kind process_document — in blocks, in instruction events and in the context watchdog's context.block.kind attribute — becomes team_process.
  • SDK. The ProcessDocument* types and schemas become TeamProcess* (TeamProcess, TeamProcessEdit, EditTeamProcessRequest, TEAM_PROCESS_MAX, TEAM_PROCESS_PATHS, …), and Client.getProcessDocument, processDocumentHistory and writeProcessDocument become getTeamProcess, teamProcessHistory and writeTeamProcess. csuite-core now exports createSqliteTeamProcessStore, TeamProcessStore and TeamProcessError.
  • Permission. process.manage becomes team_process.manage. There is no permissions migration: raw_permissions is stored verbatim and resolved on every read, and the resolver maps the old name to the new one, so a member or template saved under process.manage keeps the capability with no row rewritten. Both names are accepted on write.
  • Database. The tables process_document and process_document_edits become team_process and team_process_edits. The store renames them in place on first open, deciding from sqlite_master rather than from a caught error: a fresh database gets the new names directly, a pre-rename database is renamed with every row — including the append-only edit history — preserved, and an already-migrated database is left untouched, so opening twice is the same as opening once.

  Upgrade brokers and runners together. A runner built before this change validates the GET /instructions packet against a block-kind enum that has no team_process in it and rejects the whole response, so a new broker leaves an older runner unable to fetch its instructions at all — the same lockstep 0.9.0 asked for.

  The instruction-version hash is computed from the same inputs as before, so upgrading does not mark any runner restart-pending.

Patch Changes

• #311 20cd8c1 Thanks @andrew-jon-p7a! - Instruction and environment restarts now start the agent cold instead of resuming the prior conversation. The successor gets the refreshed instructions and the runner's context_refresh re-brief on its open objectives, and its session_start says why (resumed: false with resumeReason: 'instructions changed', 'environment changed', or 'environment reloaded'); a fresh MCP session is always re-briefed, even inside the cooldown that folds an attach and a compaction. Resuming there was the source of a capture defect measured in production: a restarted runner re-read the whole resumed transcript as new activity — 1000+ rows spanning 15 hours in one second — saturating the uploader and dropping 40–88% of events. Two fixes close that regardless of how a session restarts: the transcript reader follows a changed transcript_path (a new file after any agent swap) while keeping its line-uuid dedup, and every transcript-derived ActivityEvent carries an optional sourceId that the broker dedups on (member_activity.source_id, a partial unique index added lazily; id-less events from older runners are stored as before). The runner no longer predicts resume from the filesystem: bare --resume on csuite claude hands the decision to Claude Code (continue, which starts fresh when nothing exists), and on csuite codex starts a new thread loudly, since codex resumes by id only; explicit --resume <id> stays strict on both.

csuite-core

Minor Changes

• #311 20cd8c1 Thanks @andrew-jon-p7a! - Breaking. process_document is renamed team_process across the agent-facing surface, the wire, the SDK and storage. "Document" named the storage shape, not the concept: it framed the team's process as a static artifact to fetch rather than operational context to consult, and neither the permission leaf (process.manage) nor the store's own description had ever used the word. Bare process was not an option — it shadows Node's global and reads as the runtime process at every call site — and team_process rules that reading out on sight.

  What moves:

  • MCP tools. process_document_get, process_document_history and process_document_write become team_process_get, team_process_history and team_process_write. Only the new names are advertised in tools/list. The old names still dispatch for this one release, because tool names are quoted in prose that does not auto-update (team instructions, the team process text itself) and a hard rename would strand a caller reading them out of it; each such call logs a warn record naming the alias and its replacement so stale references can be found. The aliases are removed in the next minor.
  • REST. GET/PUT /process-document and GET /process-document/history become /team-process and /team-process/history. The GET /instructions field processDocument becomes teamProcess, and the instruction-block kind process_document — in blocks, in instruction events and in the context watchdog's context.block.kind attribute — becomes team_process.
  • SDK. The ProcessDocument* types and schemas become TeamProcess* (TeamProcess, TeamProcessEdit, EditTeamProcessRequest, TEAM_PROCESS_MAX, TEAM_PROCESS_PATHS, …), and Client.getProcessDocument, processDocumentHistory and writeProcessDocument become getTeamProcess, teamProcessHistory and writeTeamProcess. csuite-core now exports createSqliteTeamProcessStore, TeamProcessStore and TeamProcessError.
  • Permission. process.manage becomes team_process.manage. There is no permissions migration: raw_permissions is stored verbatim and resolved on every read, and the resolver maps the old name to the new one, so a member or template saved under process.manage keeps the capability with no row rewritten. Both names are accepted on write.
  • Database. The tables process_document and process_document_edits become team_process and team_process_edits. The store renames them in place on first open, deciding from sqlite_master rather than from a caught error: a fresh database gets the new names directly, a pre-rename database is renamed with every row — including the append-only edit history — preserved, and an already-migrated database is left untouched, so opening twice is the same as opening once.

  Upgrade brokers and runners together. A runner built before this change validates the GET /instructions packet against a block-kind enum that has no team_process in it and rejects the whole response, so a new broker leaves an older runner unable to fetch its instructions at all — the same lockstep 0.9.0 asked for.

  The instruction-version hash is computed from the same inputs as before, so upgrading does not mark any runner restart-pending.

Patch Changes

• #311 20cd8c1 Thanks @andrew-jon-p7a! - Instruction and environment restarts now start the agent cold instead of resuming the prior conversation. The successor gets the refreshed instructions and the runner's context_refresh re-brief on its open objectives, and its session_start says why (resumed: false with resumeReason: 'instructions changed', 'environment changed', or 'environment reloaded'); a fresh MCP session is always re-briefed, even inside the cooldown that folds an attach and a compaction. Resuming there was the source of a capture defect measured in production: a restarted runner re-read the whole resumed transcript as new activity — 1000+ rows spanning 15 hours in one second — saturating the uploader and dropping 40–88% of events. Two fixes close that regardless of how a session restarts: the transcript reader follows a changed transcript_path (a new file after any agent swap) while keeping its line-uuid dedup, and every transcript-derived ActivityEvent carries an optional sourceId that the broker dedups on (member_activity.source_id, a partial unique index added lazily; id-less events from older runners are stored as before). The runner no longer predicts resume from the filesystem: bare --resume on csuite claude hands the decision to Claude Code (continue, which starts fresh when nothing exists), and on csuite codex starts a new thread loudly, since codex resumes by id only; explicit --resume <id> stays strict on both.

csuite-sdk

Minor Changes

• #311 20cd8c1 Thanks @andrew-jon-p7a! - Breaking. process_document is renamed team_process across the agent-facing surface, the wire, the SDK and storage. "Document" named the storage shape, not the concept: it framed the team's process as a static artifact to fetch rather than operational context to consult, and neither the permission leaf (process.manage) nor the store's own description had ever used the word. Bare process was not an option — it shadows Node's global and reads as the runtime process at every call site — and team_process rules that reading out on sight.

  What moves:

  • MCP tools. process_document_get, process_document_history and process_document_write become team_process_get, team_process_history and team_process_write. Only the new names are advertised in tools/list. The old names still dispatch for this one release, because tool names are quoted in prose that does not auto-update (team instructions, the team process text itself) and a hard rename would strand a caller reading them out of it; each such call logs a warn record naming the alias and its replacement so stale references can be found. The aliases are removed in the next minor.
  • REST. GET/PUT /process-document and GET /process-document/history become /team-process and /team-process/history. The GET /instructions field processDocument becomes teamProcess, and the instruction-block kind process_document — in blocks, in instruction events and in the context watchdog's context.block.kind attribute — becomes team_process.
  • SDK. The ProcessDocument* types and schemas become TeamProcess* (TeamProcess, TeamProcessEdit, EditTeamProcessRequest, TEAM_PROCESS_MAX, TEAM_PROCESS_PATHS, …), and Client.getProcessDocument, processDocumentHistory and writeProcessDocument become getTeamProcess, teamProcessHistory and writeTeamProcess. csuite-core now exports createSqliteTeamProcessStore, TeamProcessStore and TeamProcessError.
  • Permission. process.manage becomes team_process.manage. There is no permissions migration: raw_permissions is stored verbatim and resolved on every read, and the resolver maps the old name to the new one, so a member or template saved under process.manage keeps the capability with no row rewritten. Both names are accepted on write.
  • Database. The tables process_document and process_document_edits become team_process and team_process_edits. The store renames them in place on first open, deciding from sqlite_master rather than from a caught error: a fresh database gets the new names directly, a pre-rename database is renamed with every row — including the append-only edit history — preserved, and an already-migrated database is left untouched, so opening twice is the same as opening once.

  Upgrade brokers and runners together. A runner built before this change validates the GET /instructions packet against a block-kind enum that has no team_process in it and rejects the whole response, so a new broker leaves an older runner unable to fetch its instructions at all — the same lockstep 0.9.0 asked for.

  The instruction-version hash is computed from the same inputs as before, so upgrading does not mark any runner restart-pending.

Patch Changes

• #311 20cd8c1 Thanks @andrew-jon-p7a! - Instruction and environment restarts now start the agent cold instead of resuming the prior conversation. The successor gets the refreshed instructions and the runner's context_refresh re-brief on its open objectives, and its session_start says why (resumed: false with resumeReason: 'instructions changed', 'environment changed', or 'environment reloaded'); a fresh MCP session is always re-briefed, even inside the cooldown that folds an attach and a compaction. Resuming there was the source of a capture defect measured in production: a restarted runner re-read the whole resumed transcript as new activity — 1000+ rows spanning 15 hours in one second — saturating the uploader and dropping 40–88% of events. Two fixes close that regardless of how a session restarts: the transcript reader follows a changed transcript_path (a new file after any agent swap) while keeping its line-uuid dedup, and every transcript-derived ActivityEvent carries an optional sourceId that the broker dedups on (member_activity.source_id, a partial unique index added lazily; id-less events from older runners are stored as before). The runner no longer predicts resume from the filesystem: bare --resume on csuite claude hands the decision to Claude Code (continue, which starts fresh when nothing exists), and on csuite codex starts a new thread loudly, since codex resumes by id only; explicit --resume <id> stays strict on both.

csuite-web-ui

Minor Changes

• #311 20cd8c1 Thanks @andrew-jon-p7a! - Breaking. process_document is renamed team_process across the agent-facing surface, the wire, the SDK and storage. "Document" named the storage shape, not the concept: it framed the team's process as a static artifact to fetch rather than operational context to consult, and neither the permission leaf (process.manage) nor the store's own description had ever used the word. Bare process was not an option — it shadows Node's global and reads as the runtime process at every call site — and team_process rules that reading out on sight.

  What moves:

  • MCP tools. process_document_get, process_document_history and process_document_write become team_process_get, team_process_history and team_process_write. Only the new names are advertised in tools/list. The old names still dispatch for this one release, because tool names are quoted in prose that does not auto-update (team instructions, the team process text itself) and a hard rename would strand a caller reading them out of it; each such call logs a warn record naming the alias and its replacement so stale references can be found. The aliases are removed in the next minor.
  • REST. GET/PUT /process-document and GET /process-document/history become /team-process and /team-process/history. The GET /instructions field processDocument becomes teamProcess, and the instruction-block kind process_document — in blocks, in instruction events and in the context watchdog's context.block.kind attribute — becomes team_process.
  • SDK. The ProcessDocument* types and schemas become TeamProcess* (TeamProcess, TeamProcessEdit, EditTeamProcessRequest, TEAM_PROCESS_MAX, TEAM_PROCESS_PATHS, …), and Client.getProcessDocument, processDocumentHistory and writeProcessDocument become getTeamProcess, teamProcessHistory and writeTeamProcess. csuite-core now exports createSqliteTeamProcessStore, TeamProcessStore and TeamProcessError.
  • Permission. process.manage becomes team_process.manage. There is no permissions migration: raw_permissions is stored verbatim and resolved on every read, and the resolver maps the old name to the new one, so a member or template saved under process.manage keeps the capability with no row rewritten. Both names are accepted on write.
  • Database. The tables process_document and process_document_edits become team_process and team_process_edits. The store renames them in place on first open, deciding from sqlite_master rather than from a caught error: a fresh database gets the new names directly, a pre-rename database is renamed with every row — including the append-only edit history — preserved, and an already-migrated database is left untouched, so opening twice is the same as opening once.

  Upgrade brokers and runners together. A runner built before this change validates the GET /instructions packet against a block-kind enum that has no team_process in it and rejects the whole response, so a new broker leaves an older runner unable to fetch its instructions at all — the same lockstep 0.9.0 asked for.

  The instruction-version hash is computed from the same inputs as before, so upgrading does not mark any runner restart-pending.