Decide: context management for obot-prime (C1–C6) — state file, Navigator, rehydration

[obotclaw[bot]]

Decision artifact: https://jwildfire.github.io/obot.roadmap/reports/decisions/2026-08-14-prime-context-management/ — full argument, measurements, and citations there; this thread is the place to record the calls.

The question: obot-prime stays up indefinitely, so context compaction is a certainty. What is the mechanism that keeps its answers fast AND grounded after compaction — given the <30s / ≤2-round-trip contract is the objective function? Tonight's case study: the false "monitor self-corrects" claim plus three provenance errors, all summaries-of-summaries.

Six calls (answer one by one):

C1 — State vs narrative. Give prime a capped, provenance-stamped prime-state.md (Open / Delegates / Armed / Claims / Settled; every line stamped [verified gh HH:MM] vs [asserted by X]; resolved lines deleted)? → Recommend yes (options: a status quo / b state file / c JSON).

C2 — The Navigator (your mid-session proposal). → Recommend yes, strictly as a file-writing verifier — never someone prime asks (a peer prime must ask is a round trip and busts the 30s contract; a peer that writes a file prime reads is strictly better). Day one: bookkeeping/verification only; standards enforcement as observe-and-report, never correcting other agents; audits stay in their existing lane for now. Ramp from a scheduled sweep (tonight's 22:51 sweep is the prototype — it caught the stale oa#83 claim) to a standing session only if cadence demands.

C3 — Rehydration. One scripted, one-round-trip bundle read (prime-rehydrate: state + today's log + sweep cache ≈ 7k tokens, 0.35s measured), fired lazily on the first post-compaction turn that touches state? → Recommend yes (options: a re-read whole log / b scripted bundle / c trust the summary).

C4 — Write cost. Replace the 759-char heartbeat heredoc (≈190 generated tokens ≈ 2–3s per event) with a scratchpad-log script (≈20 tokens, ~10× cheaper, benefits every sibling)? → Recommend yes.

C5 — Letting go. Adopt four retention rules (answered Q&A dropped once relayed; delegate reports → one pointer; GitHub state pointed at, never copied; settled decisions = one line until applied, then deleted)? → Recommend yes.

C6 — Where it lands. One obot.agent PR (skill section + framework link-section + two scripts + template edit) left open for your review, and the Navigator as a hub requirement issue through the normal lifecycle? → Recommend yes.

---

This discussion was drafted by Claude Code using Fable 5 in an unattended session and has not been reviewed by @jwildfire.

[obotclaw[bot]]

Decided 2026-08-14 — approved. @jwildfire (relayed via 🎩🤖 obot-prime): "I'm good with your recommendations in https://jwildfire.github.io/obot.roadmap/reports/decisions/2026-08-14-prime-context-management/"

Implemented 2026-08-15, day-one phases only:

• C1/C3/C4/C5/C6 — obot.agent#91, merged on the standard lane (closes oa#90): session-prime skill gains the durable-state contract; tools/prime-rehydrate (measured 25.4 KB / 0.02 s tonight, byte-bounded log section) pre-injected by /s-prime; tools/scratchpad-log with the heredoc kept as template fallback so in-flight siblings are unaffected; link-only standing-sessions section in session-framework.md.
• C2 (Navigator) — filed as requirement #157 under goal Goal: increased autonomy in obot.agent #73, milestone 2026q3: file-writing verifier only, day-one bookkeeping scope, observe-and-report standards, audits stay put.
• Prime's prime-state.md was seeded empty; prime starts writing it this session.

Two deviations from the artifact's C6 text, per @jwildfire's rules set 2026-08-14: the PR was self-merged on the standard lane rather than left open (only RC PRs reach the review queue), and the Navigator issue carries milestone 2026q3.

Still @jwildfire's edits to make: one workspace-settings allowlist line for scratchpad-log; any # Compact instructions CLAUDE.md block was deliberately not added (workspace-wide effect — say the word if wanted).

---

This comment was drafted by Claude Code using Fable 5 in an unattended session.