v0.10.0 — Source-bound project knowledge

Pallium now gives agents compact project context with visible freshness, source evidence, and audit limits.

What changed

• Workspace-specific indexes and source-bound freshness prevent old prose from appearing current after edits.
• Exact module and receiver citations, collision-safe module IDs, improved TypeScript scanning, and explicit parser limitations.
• Compact CLI/MCP search and task context, source sections, tracked documentation/configuration, and explicitly linked decisions.
• A local knowledge browser with source evidence, audit outcomes, light/dark themes, and maintenance controls.
• Opt-in background upkeep with durable recovery, a global 40-attempt rolling daily cap, two active model slots, and deadlines.

Upgrade

npm install -g pallium@0.10.0
pallium index
pallium knowledge build --no-model

Legacy documents remain readable with rebuild labels. Compact response envelopes are version 2; use --full when full document payloads are needed. Use returned module IDs instead of constructing old slugs. Index linked worktrees separately.

Enable maintenance only for repositories whose source you intend the configured provider to read. Large initial builds remain subject to the same daily cap.

Evidence and limitations

All 30 controlled Go/TypeScript/Swift retrieval questions found the expected module in the top three. Two paired agent tasks used about 96% less supplied context with correct answers in both variants; no latency improvement or broad token-cost savings are claimed. Generated claims remain fallible, and heuristic parsers are incomplete.

See docs/releases/v0.10.0.md for migration and rollback details, and eval/knowledge/README.md for evaluation methodology.

PR: #85

v0.9.23 — Update itself

v0.9.23 — Update itself

pallium update brings the CLI current from wherever it is installed. No more
remembering whether a machine got pallium through npm, a release tarball, or
go install — the command detects the install shape and does the right thing.

What's new

• Self-update. pallium update [--json] checks the latest GitHub release
  against the running build and upgrades it. npm-managed installs (the binary
  under ~/.pallium/npm/) go through npm install -g pallium@latest so the
  package metadata stays truthful; a bare binary gets the verified release
  asset swapped in place — same checksums.txt verification the installer
  uses, staged next to the executable and renamed atomically.
• Honest fallbacks. No npm on PATH under an npm layout falls back to the
  binary swap; platforms without prebuilt assets say so and point at go install; a checksum mismatch leaves the existing binary untouched.
• Devin, first-class. This release also carries the Devin provider and
  session support shipped in v0.9.22 for anyone updating from earlier.

Install

npm install -g pallium
pallium version

Then keep it current with:

pallium update

v0.9.22 — Devin, first-class

v0.9.22 — Devin, first-class

Pallium now treats the Devin CLI like Codex and Claude Code: workflows can
run Devin-backed workers, teams can give Devin teammates persistent native
sessions, and session awareness sees Devin CLI sessions alongside the others.
Run pallium from inside a Devin session and it adopts Devin automatically —
no configuration.

What's new

• Devin as a workflow provider. agent(..., { provider: "devin" }) runs
  the devin CLI headless, capturing the final answer, session id, and token
  counts from the ATIF export rather than stdout. Read-only agents run
  --permission-mode auto (fail-closed: write, exec, and networked tool calls
  are all rejected); edit/test/check agents run dangerous. Steering
  detection recognizes Devin via CHISEL_SESSION_DB, and
  PALLIUM_WORKFLOW_PROVIDER_DEVIN_COMMAND still overrides with a wrapper.
• Devin teammates with real session continuity. Team members on Devin
  resume their own native session across turns via the exported session id —
  the same contract codex and claude teammates already had.
• Devin in pallium sessions. sessions live, watch, and find read
  Devin's session database and classify lifecycle from message history —
  running, waiting, blocked, finished — alongside Codex and Claude.
• Devin session memory. sessions index/sync --provider devin
  (--devin-db for a non-default database) normalize Devin's SQLite history
  into the shared schema, so search, recall, show, and read answer
  from Devin transcripts with devin: citations.

Known limitations

• Devin edit agents may reach the network regardless of network: true.
  --permission-mode dangerous auto-approves networked tools; conversely a
  networked read-only request can't be granted (auto blocks egress) and
  logs a warning. Use a configured wrapper for tighter egress control.
• Devin cost is not tracked in USD. Devin reports ACU/token metrics, so
  it joins codex on the untracked-cost list — team status/loop status
  say so explicitly instead of showing an indistinguishable $0.0000.

Install

npm install -g pallium
pallium version

v0.9.21 — Know the codebase, not just its history

v0.9.21 — Know the codebase, not just its history

Pallium's repo intelligence was built entirely from git history. It could tell
you which files move together and how risky one is to touch, and nothing at all
about what any of them contained. There was no symbol anywhere in the index and
no command that searched code.

This release adds the other half: a content index of symbols and import edges,
and a knowledge base built on top of it that refuses to store a claim it cannot
check.

What's new

• pallium index indexes content, not just commits. Per file it records a
  content hash, its symbols with kind, line range, signature and doc comment,
  its resolved import edges, and the identifiers it references. Go parses
  through go/parser; TypeScript, JavaScript, Python, Swift, Rust, Ruby, Java,
  Kotlin, C#, PHP and C/C++ use line-anchored scanners, because Pallium ships
  as a single cgo-free binary and a native grammar would break both
  cross-compilation and the npm install. The content half is keyed by hash, so
  reindexing only reparses what changed: 2,458 symbols across 186 files in
  0.43s on Pallium itself, and a second run reparses nothing.

• pallium symbols <path> lists what a file declares, with kinds, line
  numbers, signatures and doc comments, plus the files it imports and the files
  that import it.

• pallium callers <name> separates where a name is declared from where it
  is merely referenced, which is what sizing a rename actually needs.

• pallium search <query> runs FTS5 over symbol names, signatures and doc
  comments, name matches ranked first.

• pallium knowledge map|build|audit|list|get|search builds a readable
  description of what each part of a repo is for. Modules are clustered from
  the content index; their file lists, symbol surfaces ranked by how many files
  reference each name, dependency edges in both directions, external packages
  and recent history are all derived, so they are correct by construction. An
  incidents doc is mined from commit subjects announcing reverts, rollbacks,
  hotfixes and outages, so every entry is a real SHA with a real file list.
  --no-model builds the whole base with no model calls at all.

• A model fills a schema; it never writes the page. Synthesis returns a
  fixed JSON shape of cited claims and the renderer builds the markdown, so
  there is no path for an unchecked citation to reach a doc body. Every cited
  path and symbol is checked against the index before storage; a claim citing
  something that does not exist is deleted, recorded in the doc's dropped
  claims, and the doc is marked unverified.

• pallium knowledge audit checks whether claims are true, not just
  citable. Verification proves a citation resolves. It does not prove the
  claim about that symbol is correct: "Open deletes the database", citing a
  real Open, passes. The audit hands a skeptic each stored claim together
  with the real source of every symbol it cites and removes what the source
  does not support.

  Measured on Pallium's own knowledge base: 325 claims, 134 supported, 168
  unclear, 23 removed. Four removals were spot-checked against the source by
  hand and all four were correct — the audit caught a claim that the indexer
  wraps branch detection inside its transaction (it does not; two git reads
  happen before it opens), a claim that a recency clamp guards against
  same-day commits (the expression already adds one, so the clamp only fires
  on future timestamps), a claim that a reindex wipes every repo table (the
  delete list is explicit), and an invented eight-value symbol-kind
  vocabulary that nothing enforces.

  The unclear rate is high and that is the intended behavior, not a
  shortfall: over half of all claims concern behavior spanning more code than
  one declaration's excerpt shows, and unclear is a first-class verdict
  because forcing a binary answer on insufficient evidence either deletes good
  knowledge or keeps bad. The audit only ever removes what it can justify.

• Seven knowledge tools on the MCP surface. pallium workflow mcp
  previously exposed workflow tools only, so a connected agent could start a
  run but could not ask one question about the repo. It now serves
  pallium_knowledge_search, _get and _map, plus pallium_explain,
  _symbols, _callers and _search_code. Knowledge search is BM25-ranked,
  not substring matching.

• pallium explain says what a file is before saying how risky it is. Its
  summary used to be one of three fixed strings chosen by risk level. It now
  leads with what the file declares and which module it belongs to.

• Synthesis fans out. One provider call per module measured at ~55s, so
  knowledge build runs them concurrently (--concurrency, default 4) while
  verification and storage stay strictly sequential, since a repo's sqlite file
  has exactly one writer by design. Pallium's own 18 modules build in 2m55s.

Fixed

• explain, risk and review no longer read your dependency tree to
  answer one question. The file listing walked the working tree and skipped
  only .git, .pallium and .codex-memory, so a repo with node_modules or
  vendor paid for all of it on every call. Measured on a three-file repo with
  one real import: 1.9s clean, 12.2s with 20,000 gitignored files beside it,
  same single result. It now asks git ls-files, which inherits .gitignore,
  and caps scanned files at 512KB so a minified bundle is never regexed.
• TypeScript ESM import specifiers resolve. An import written
  from "../server/auth.js" whose source on disk is auth.ts is TypeScript's
  NodeNext convention, not an oddity. It resolved to nothing, which meant an
  entire directory of a real 858-file Next.js app reported zero in-repo
  dependencies.
• A parser improvement now reaches existing indexes. A content hash cannot
  notice that Pallium got better at reading a file, so after an upgrade every
  hash matched, nothing reparsed, and a stale index persisted indefinitely.
  code_files stores a versioned parser tag and a mismatch forces a reparse.
• Import resolution has one implementation. internal/analysis carried its
  own copy of the Go, tsconfig-alias and Python resolution rules, which had to
  stay in sync with the indexer by hand. Both paths now call
  codeindex.Parse, and a test pins the indexed and fallback implementations
  of StructuralLinks to byte-identical output so the fallback cannot rot.
• An audit cannot report unauditable docs as clean. Docs written before
  claims were stored alongside the rendered body have nothing to re-check;
  they are now counted separately and named, instead of being folded into
  "skipped as structural-only".
• Evidence is scoped to the module being documented. Symbol names repeat
  across packages — Pallium declares Store in four, Run in six — so
  resolving a cited name repo-wide and taking the first match handed the
  auditor a stranger's source. It then refuted 44 true claims, correctly, on
  code that had nothing to do with them. Resolution now prefers the claim's
  own cited paths, then the module's files, then reports the symbol as not
  found, because silently omitting evidence makes absence of proof read as
  disproof.
• Neither build nor audit will touch a module the index has not seen. Both
  hand stored line numbers, signatures and doc comments to a model as fact, so
  a file that grew since indexing points them at whatever now occupies those
  lines. One file had drifted 28 lines and an excerpt labelled *Store.Repo
  was actually WithTx's body. A module whose files no longer hash to the
  index is skipped and counted (--allow-stale overrides on build), and every
  excerpt must show its declaration's own name within three lines of where the
  index claims it starts, so any future drift degrades to missing evidence
  rather than wrong evidence.

Upgrading

Existing indexes keep working: a repo without a content index falls back to the
previous behavior, and pallium index fills it in. Run pallium index once per
repo after upgrading, then pallium knowledge build --no-model for a base that
costs nothing.

Install

npm install -g pallium
pallium version

v0.9.20 — Route model and effort deliberately

v0.9.20 — Route model and effort deliberately

Pallium can now select provider, model, and reasoning effort through one
inspectable policy. Explicit choices still win, Auto routing remains opt-in,
and every provider attempt records the configuration and available usage
evidence needed to evaluate routing safely.

What's new

• Route complete execution configurations. pallium route models manages
  provider, model, and reasoning-effort choices together instead of treating
  the model name as the whole decision.
• Apply routing across workflows and teams. Agents, checks, gates, and team
  members share the same policy while preserving explicit pins and allowed
  provider boundaries.
• Inspect invocation evidence. Workflow reports record requested execution
  settings, duration, status, token usage, and known cost across retries.
• Evaluate policies locally. A versioned fixture set and evaluation harness
  support bounded trials, independent grading, comparison reports, and shadow
  policy proposals without automatically promoting an unproven route.

Fixed

• Evaluation results cannot mix different execution identities. Candidate,
  routing-config, Pallium binary, and worker CLI identities are pinned before
  results can be compared or proposed.
• Team Auto routing can choose the provider. An omitted provider remains
  unpinned until the routing policy makes its selection.
• Objective checks include untracked files. Evaluation trials now reject
  newly created test files as well as modified tracked tests.
• Invocation history distinguishes configuration failures. Unsupported or
  unavailable configurations are recorded as not dispatched instead of being
  reported as sent to a provider.
• Gate approvals follow the effective route. Cached approvals are
  invalidated when provider availability or another routing input changes the
  selected provider, model, or effort.
• Claude failure accounting keeps reported usage. Error envelopes retain
  token and cost evidence so failed paid attempts remain inside workflow and
  team budget accounting.
• Codex event streams stay memory-bounded. Usage is accumulated while the
  stream is read, and only a small diagnostic tail is retained for failures.
• Real routing trials cannot inherit drifting Codex settings. Non-simulation
  evaluation runs require the isolated Codex configuration path.
• Team policy checks use the live repository root. Edit worktrees can no
  longer miss a gitignored routing policy after an operator removes a provider.
• Team routing covers the complete member lifecycle. Task classes reach the
  policy, Codex executable checks are deferred to the configured team runner,
  plan-required members select an edit-eligible route, and pre-dispatch team
  failures are recorded honestly.
• Routing evidence stays attached to its real source. Policy suggestions
  require the same recorded config, malformed or empty invocation snapshots
  cannot be graded, redirected test databases remain isolated, and
  workflow-created team gates retain invocation history.
• A dispatch begins only when a provider process starts. Validation and
  process-start failures from workflows and teams are retained as
  not_dispatched instead of disappearing or posing as executed calls.
• Provider capabilities match the actual adapter. Configured wrappers own
  their reasoning-effort vocabulary, while the networkless built-in Claude
  adapter is excluded from routes that require network access.

Install

npm install -g pallium
pallium version

v0.9.19 — Use the agent, not the menu

v0.9.19 — Use the agent, not the menu

Pallium now helps an agent choose and use the right capability from a plain
statement of intent. It also understands session completion and recency, so an
agent can reason about current and recent work without treating transcript
keywords or process names as proof.

What's new

• Route and act from intent. pallium route <task> --execute selects a named
  capability, explains its evidence and alternatives, and invokes the command
  with structured arguments when it fits the caller's existing authority.
• Discover the decision contract. pallium route capabilities --json
  exposes when each capability fits, when to avoid it, its required authority,
  and the evidence that demonstrates success.
• Ask natural questions about sessions. pallium sessions find <query>
  understands recently finished, unfinished, inactive, and most-recently
  updated sessions. It returns the interpretation and filters it applied.
• Distinguish runtime state from completion. Session results separately
  report whether a process is active and whether transcript evidence shows the
  task finished, did not finish, or remains unknown.

Fixed

• Live discovery no longer counts helper processes as agent sessions. Exact
  process identity prevents Codex host helpers from appearing as duplicate
  running work.
• Live-state claims carry evidence and coverage. State reasons, sources,
  confidence, timestamps, provider coverage, and explicit exclusions prevent a
  best-effort local view from posing as an exhaustive computer inventory.
• Running-session requests exclude completed open tasks. The router uses
  sessions live --running-only, keeping finished desktop tabs out of the
  answer while retaining their lifecycle evidence when explicitly requested.
• Workflow concurrency tests no longer depend on fixed sleeps. Marker-based
  synchronization removes a timing-sensitive failure from the full test suite.

Install

npm install -g pallium
pallium version

v0.9.18 — Pick up where you left off

v0.9.18 — Pick up where you left off

Pallium can now turn prior Codex and Claude sessions into useful continuity:
what the work was, where it stopped, and what should happen next. This release
also restores Pallium's optional MCP tools in Codex by speaking the standard
stdio transport correctly.

What's new

• Recall work across sessions. pallium sessions recall <question> combines
  lexical and semantic retrieval, then returns a bounded continuity capsule
  with the goal, decisions, files, blockers, and next action from the best
  matching session.
• Keep the session index healthy. Session sync, maintenance, and retrieval
  now cover Codex and Claude history more reliably, including archived sessions,
  legacy records, duplicate messages, stale embeddings, and permanently failed
  embedding jobs.
• Configure embeddings once. Embedding settings persist locally, with API
  credentials stored in the macOS Keychain instead of the Pallium database.
• Find a Codex goal file directly. pallium sessions goal <session-id>
  resolves the goal attachment for a session. Use --path-only for scripts or
  --json for structured output.

Fixed

• Pallium MCP works in Codex again. The workflow MCP server now uses the
  newline-delimited JSON transport required for stdio MCP servers, eliminating
  the startup timeout caused by the old Content-Length framing.
• Incomplete workflow schemas fail early and clearly. Startup now verifies
  that existing workflow, team, and loop tables contain every expected column,
  so a missed migration is reported immediately instead of surfacing later as
  an unrelated query failure.

Install

npm install -g pallium
pallium version

v0.9.17 — Teams, made adoptable

v0.9.17 — Teams, made adoptable

This release closes the teams trilogy: possible, composable, and now
adoptable. Ready-made team shapes, a way for any running agent session to
join a team directly, and a more forgiving decision format.

What's new

• Team templates. pallium team start <goal> --template parallel-review
  or adversarial-debate spawns a known-good team shape in one command —
  distinct-lens reviewers on the same artifact, or two members arguing
  opposite sides of a question. team template list to browse.
• Join a team from any session. pallium team join <team-id> --as <name>
  lets an already-running agent session — another editor tab, a different
  CLI, a person at a terminal — join a team and coordinate through its
  shared mailbox, with no separate setup.
• A more forgiving decision format. Teammates no longer need to include
  every optional field for their decisions to count. A genuinely broken
  decision is never silently ignored either — you'll hear about it.
• The setup guide now covers teams and loops. The trigger Pallium
  installs into your project only ever mentioned workflows before this
  release. It now mentions all of Pallium's capabilities, so a fresh agent
  session can actually discover team and loop support.

Fixed

• Task completion is now reliable. A teammate that finished real work
  could previously leave the task board showing it as still pending. Fixed.

Install

npm install -g pallium
pallium version

v0.9.16 — Teams become composable

v0.9.16 — Teams become composable

Agent Teams can now be driven from workflow scripts, need real plan review
before editing, and support quality gates and per-teammate supervision.

What's new

• Teams from workflow scripts. A workflow (or a loop tick) can now
  create and drive a whole team programmatically — spawn teammates, send
  messages, wait for convergence, and read back the result, all from code.
• Plan approval. A teammate can be required to submit a plan and wait for
  it to be approved before it's allowed to make any edits.
• Quality gates. Configure an autonomous verifier that checks a team's
  work at key points — when a task is created, when one is completed, or
  when a teammate goes idle — and can send it back for another pass.
• Supervise one teammate at a time. team member stop|restart|steer lets
  you pause, resume, or redirect a single teammate without touching the rest
  of the team. Changes take effect at that teammate's next turn, so nothing
  is ever interrupted mid-thought.
• Honest cost reporting. Team spend reporting now says plainly which
  providers don't report real usage data, instead of showing a misleading
  $0.00.

Install

npm install -g pallium
pallium version

v0.9.15 — Operational honesty

v0.9.15 — Operational honesty

Pallium now tells you plainly when something is actually done, and cleans up
after itself when a process dies unexpectedly.

What's new

• Stale runs clean themselves up. If the process running a workflow gets
  killed, its run and workers used to look "running" forever. Every run now
  tracks its own liveness and automatically flips to an honest "interrupted"
  status once its owning process is gone — checked every time you list,
  inspect, or check on your runs. workflow gc --stale runs the same cleanup
  as an explicit command.
• An unmissable verdict. workflow status and report now lead with a
  blunt "FINISHED" or "NOT FINISHED" line, so it's never ambiguous whether a
  run actually completed.
• The setup prompt installs itself. pallium start now offers to add
  Pallium's adoption block to your AGENTS.md/CLAUDE.md the first time it
  notices one is missing, instead of leaving that step for you to remember.

Install

npm install -g pallium
pallium version