PR #1199 — Grounding Charters in the Corpus, Not the Spec

The Broader Goal at the End of the Previous PR

PR #1202 (post #132) finished a mechanical terminology sweep — replacing stray "agent" with "harness" across synlynk/*.py so the Agent/Harness distinction (agents are who, harnesses are how) held consistently in code, not just in docs. That PR closed one gap on tracking issue #1198 ("[Tracking] Autonomous Operations Activation"): the terminology was now internally consistent, but several sub-issues under #1198 remained — including #1199, which #1202's own investigation had flagged as still open. The goal at that point was simply: work through #1198's remaining sub-issues in priority order, starting with whichever was shovel-ready.

Strategic Shifts in This PR

None. #1199 was picked over #1203 (GOVERNS backlog automation) specifically because #1203 requires a superpowers:brainstorming pass first per its own issue text, while #1199 — "Document corpus references used to derive workspace agent charter content" — could go straight to a plan. No goalpost moved; this was the next item in the queue, taken in the order its own preconditions allowed.

What This PR Shipped

The real discovery this PR made was architectural, not textual: charter content lives in two places, and #1199 only owns one of them.

1. synlynk/agent_cli.py's SEED_CHARTERS dict — the git-tracked, durable seed source used only when cmd_agent_init() provisions a brand-new agent.
2. Each live agent's actual runtime charter at ~/.synlynk/workspaces/<workspace_id>/agents/<agent_id>/charter.md, written by agent_store.propose_charter_revision() — not git-tracked, workspace-local state.

PR #1196 had already restored pm's charter text from a real record; #1199's job was to do the same derivation exercise — for real, from the actual corpus, not spec-guessed prose — for the other six roles: dev, qa, architect, tpm, designer, marketing. synlynk-bot was explicitly out of scope (infra catch-all, no real corpus to mine).

Per this repo's locked role split, the actual charter authorship was dispatched, not written by Claude:

• Task 1 (dispatched to Codex, job job-8a37f307): audited dev, qa, architect against project-docs/devlogs/nikhilsoman.md, project-docs/memory.md, and targeted git log --oneline --grep= searches. Result: qa's and architect's charter bodies were rewritten — qa to describe policy-gated (.synlynk/policy.json), currently docs-only-demonstrated merge authority rather than an unrestricted claim; architect to describe the role as provisioned-but-largely-unexercised, since the corpus shows Claude actually operating in a pm/reviewer capacity, not a separately exercised architect identity. dev's charter was left unchanged — the corpus confirmed the existing text was already accurate. "No change" was an explicitly valid finding, not a failure to find something.
• Task 2 (dispatched to Codex, job job-f4c16542): same treatment for tpm, designer, marketing. tpm's charter now precisely describes the implemented ready-story-scan / policy-gate-check / ticket-create-or-reuse / daemon-dispatch loop (synlynk/tpm_sweep.py), citing the 2026-08-24 "ticket-driven approval auto-resume" devlog entry, rather than the vaguer prior framing. marketing's Instructions dropped an unsupported universal claim ("writes every PR's blog post") in favor of dispatch-conditional framing. designer was left unchanged — no fabricated track record found.

Every changed section only touched charter body prose — Instructions, Authority & Escalation, Workflow Ownership. Frontmatter (schema_version, role, description, durability, tools, credentials) was verified byte-identical to origin/main for all 8 roles, including the 6 touched ones, by the spec-compliance review pass.

docs/charters/corpus-references.md documents the sourcing for all 6 roles — per-role Sources consulted / Findings / Charter changes made, with specific devlog dates, memory section names, and grep commands, all independently verified to exist against the real corpus (not just asserted) by the review passes.

Both required review stages ran: a spec-compliance pass confirmed frontmatter integrity and citation accuracy; a code-quality pass confirmed internal consistency across sibling charters (no conflicting merge-authority claims between qa/architect/tpm), Python string-literal correctness in the SEED_CHARTERS dict, a clean charter_schema.validate_charter() run for all 6 changed roles, and a clean full-suite test run — 2322 passed, 2 skipped, 2 pre-existing database is locked failures (unrelated to this change, present on origin/main too).

Merge note: the same underlying job content (job-8a37f307 + job-f4c16542) landed via two parallel PM tracks working the same issue — this session's own PR (#1262, opened after rebase/two- stage review) and a separate session's PR (#1260), which reached the daemon's "reviewer unavailable → escalate to Claude" fallback first and self-merged with an equivalent verification checklist (diff scope, charter_schema.validate_charter(), targeted pytest). #1260 is the PR of record; #1262 was closed as superseded once a rebase confirmed zero remaining diff against main beyond this blog post. Not a process gap — both tracks independently converged on the same sanctioned escalation path.

Brainstorm Visuals Used

None — #1199 had no brainstorm gate; it went straight from issue to plan per the Design → Plan → Build sequence.

What This Achieved on the Path to Autonomy

A charter that claims authority the corpus doesn't back is worse than no charter — it tells a dispatched agent (or a human reading the roster) that a role can do something it has never actually been exercised to do. Grounding six charters in real devlog/git history closes that trust gap ahead of any push toward more autonomous, charter-driven dispatch: the SEED_CHARTERS seed text a brand-new agent gets on synlynk agent init now matches what the role has actually demonstrated, not what the original spec guessed it might.

Strategic Note: The Goal at the End of This PR

Task 3 — bumping the 6 already-provisioned live agents' runtime charters (~/.synlynk/workspaces/.../charter.md) to match the updated SEED_CHARTERS text via propose_charter_revision() — is an explicit post-merge operational step performed directly by Claude (PM/deploy role), not dispatched, mirroring how PR #1196 handled the same step for pm. Once that's done, the next and final open item under #1198 is #1203 (GOVERNS backlog automation), which still needs its own superpowers:brainstorming pass before a plan can be written.