Activity

1. github-project-automation

   added this to fullsend-ai Workstreamon Oct 15, 2026
2. github-project-automation

   moved this to Todo in fullsend-ai Workstreamon Oct 15, 2026
3. fullsend-ai-triage commented on Oct 15, 2026

   

   fullsend-ai-triageBot

   on Oct 15, 2026 – with fullsend-ai-triage

   Last edited by fullsend-ai-triage

   More actions

   🤖 Finished Triage · ✅ Success · Started 4:27 PM UTC · Completed 4:29 PM UTC

   Commit: 99417cc · View workflow run →

   Runtime: claude · Model: sonnet → claude-sonnet-5-5 · Effort: high · Cost: $0.21
4. fullsend-ai-triage

   added

   featureFeature-category issue awaiting human prioritizationFeature-category issue awaiting human prioritization

   triagedTriaged but awaiting human prioritizationTriaged but awaiting human prioritization

   on Oct 15, 2026
5. fullsend-ai-triage commented on Oct 15, 2026

   

   fullsend-ai-triageBot

   on Oct 15, 2026 – with fullsend-ai-triage

   More actions

   This reads as a well-scoped feature proposal, and the code references check out against the current tree: reservedSandboxKeys carries the sync NOTE, bootstrapEnv exists, and EnvExports() is implemented by the claude, codex and pi runtimes (opencode and the dummy runtime return nil). No open PR or sub-issue addresses this, and no duplicate was found. #7010 (reservedSandboxKeys only guards env.sandbox) is related but not blocking.

   Notes for the implementer:

   • bootstrapEnv builds many exports as inline lines. To make the single-source-of-truth claim real, the always/conditional keys and their gating conditions probably need to be declared as data that both bootstrapEnv and the generator consume. A generator that re-lists the keys would just be a fourth hand-synced copy.
   • docs/contracts/ does not exist yet, so it is a new path. Check the docs-site sidebar config if the manifest or its description should be surfaced there.
   • fullsend_version in the committed manifest would make the golden test fail on every release unless it is stamped at release time or excluded from the diff.
   • The latest ADR in the repo is 0141. The proposed 0142 looks free, but confirm there is no collision before merging.
   • Per AGENTS.md, add or update docs and keep patch coverage at or above the Codecov gate.

   Proposed test case:

   TestSandboxEnvManifest_Golden: generate manifest in-process, compare to committed JSON, fail with regenerate hint on diff.
   TestBootstrapFullsendKeysAreReserved: every FULLSEND_* key exported by bootstrapEnv is in reservedSandboxKeys.
   Edge cases: nil EnvExports (opencode), conditional keys with gating text, deterministic ordering.
   

   The agents-repo change that reads the released-tag manifest is tracked separately and is not a prerequisite for the fullsend-side work. Because this is a feature, it is not eligible for automatic promotion to ready-to-code, so a maintainer should review the design (manifest path, ADR) first.

   ---

   Next steps:

   • /fs-code — agent creates a PR to implement this issue
   • /fs-code <your instruction> — agent implements with your specific guidance
6. jflowers

   added

   enhancement

   on Oct 15, 2026
7. jflowers commented on Oct 15, 2026

   

   jflowers

   on Oct 15, 2026

   ContributorAuthor

   More actions

   Divisor review panel triage: #8465

   Seven reviewers assessed this issue. All seven read it as a valid, well-scoped feature proposal, with code references verified against the current tree. The consolidated verdict is NEEDS-CLARIFICATION. The proposal looks solid overall; a few design details are worth settling before implementation begins.

   Questions from the panel:

   1. fullsend_version in the manifest. The golden test compares the committed manifest against generator output. What happens to that comparison on every release bump? Worth deciding whether the field is excluded from the diff or stamped at release time, so the gate does not fail on each release.

   2. A fifth export site. The proposal names four sources: bootstrapEnv, reservedSandboxKeys, per-runtime EnvExports(), and gcp-vertex.env host-file injection. iterationEnvCommand (internal/cli/run.go:3507) also writes FULLSEND_TIMEOUT_MINUTES, FULLSEND_ITERATION_DEADLINE, and TRACEPARENT to iteration.env, which the sandbox .env sourcing chain picks up. The example manifest lists TRACEPARENT under always but omits the two timing vars, which are both exported and reserved. Would a generator scoped to the four named sources be born incomplete?

   3. Dynamic reserved keys. init() merges oidcDenyKeys and providerOnlyKeys into reservedSandboxKeys at load time, and the map contains the computed runnerPathVar key. Should the generator consume the merged runtime state rather than the static map literal, so dynamic entries survive?

   4. Runtime-implementer scope. dummy.go and dummy_playback.go also implement EnvExports(), returning nil. Does the "all implementers represented" criterion include the test runtimes?

   5. Pi's full export set. PiRuntime exports six variables: PI_CODING_AGENT_DIR, PI_CODING_AGENT_SESSION_DIR, PI_OFFLINE, PI_SKIP_VERSION_CHECK, PI_TELEMETRY, and JITI_FS_CACHE. The example manifest shows only the first.

   6. New docs path. docs/contracts/ does not exist yet. Per AGENTS.md, the implementer should check the docs/.vitepress/config.ts sidebar config and add the folder prefix to search.options.scopes.

   Duplicate check: the search surfaced #303 (agent permissions migration), which the panel judged topically unrelated. #7010 (reservedSandboxKeys guards env.sandbox only) is related but distinct, as the earlier triage noted.

   Classification: feature (unanimous), objective, no split recommended, no duplicate. ADR 0142 is free; the latest ADR is 0141.

   This triage was performed by the Divisor review panel.
8. jflowers commented on Oct 16, 2026

   

   jflowers

   on Oct 16, 2026

   ContributorAuthor

   More actions

   Answers to the panel's questions, verified against the current tree:

   1. fullsend_version in the manifest. Dropping the field. The version variable is "dev" unless ldflags stamp it at release build time (internal/cli/root.go:12), so a golden test can never generate a real value — committed output would always read "dev". Stamping at release time conflicts with the fetch mechanism: the manifest is pinned at tags/<version>, the tag tree is frozen at tag time, and there is no pre-tag version-bump commit to carry the stamp (the version is ldflags-injected by GoReleaser, not a committed file). The consumer already knows the version by construction — it is the tag in the URL it pinned. schema_version stays for shape.

   2. Fifth export site. Correct — iterationEnvCommand (internal/cli/run.go:3506) joins the generator's source list as a fifth site, alongside bootstrapEnv, the reservedSandboxKeys merges, per-runtime EnvExports(), and gcp-vertex.env. It writes FULLSEND_TIMEOUT_MINUTES, FULLSEND_ITERATION_DEADLINE, and TRACEPARENT to iteration.env, which the sandbox .env chain sources. All three were already reservation-checked (run_test.go:4522-4524); the export side was the gap. The example manifest's always list showing TRACEPARENT without the two timing vars was this same incompleteness — the generator derives the full set from the template production writes.

   3. Dynamic reserved keys. The generator consumes the merged runtime state. init() merges oidcDenyKeys and providerOnlyKeys into reservedSandboxKeys at load (run.go:3270-3293), and runnerPathVar is already in the map literal (run.go:3264). Running the generator in-process (a Go test in internal/cli) means init() has already executed — reading the map yields merged state with no re-derivation. Re-deriving the union from the literals would duplicate the merge logic, which is the hand-sync failure mode this proposal exists to eliminate.

   4. Runtime-implementer scope. The criterion changes from "all implementers" to the config-selectable set: the manifest's runtime section derives from config.ValidRuntimes() (internal/config/config.go:300) — claude, pi, codex, dummy, dummy-playback. The dummy runtimes are config-selectable for behaviour test orgs (registry.go:20-27) and return nil from EnvExports(); opencode also returns nil but is a registered stub excluded from ValidRuntimes (registry.go:85-87) — it cannot be activated through config, so the original criterion had the two backwards. Nil-returners appear explicitly (e.g. "dummy": []) so consumers can distinguish "exports nothing" from "not represented", and the golden test asserts the manifest's runtime set equals ValidRuntimes(), so a newly added runtime fails the test until represented.

   5. Pi's full export set. Confirmed — PiRuntime.EnvExports() (internal/runtime/pi.go:145-153) exports six variables: PI_CODING_AGENT_DIR, PI_CODING_AGENT_SESSION_DIR, PI_OFFLINE, PI_SKIP_VERSION_CHECK, PI_TELEMETRY, JITI_FS_CACHE. The example manifest showed only the first for brevity; the generator parses EnvExports() output directly, so the full set appears automatically. The manifest will annotate that JITI_FS_CACHE is jiti's variable pinned by the runner, not one of pi's own, so the list is not misread.

   6. New docs path. Keeping docs/contracts/. The JSON itself needs no sidebar entry (the sidebar lists .md pages) and raw.githubusercontent.com serves it regardless of the site build. The change adds a human-readable docs/contracts/README.md describing the contract, wires the section via getMarkdownFiles("contracts", "contracts") in docs/.vitepress/config.ts, and adds "/docs/contracts/" to the Design Docs search scope prefixes (config.ts:469-482).
9. github-project-automation

   moved this from Todo to In progress in fullsend-ai Workstreamon Oct 21, 2026
10. jflowers

    self-assigned this

    on Oct 21, 2026
11. rh-hemartin commented on Oct 22, 2026

    

    rh-hemartin

    on Oct 22, 2026

    Member

    More actions

    Adding, renaming, or reserving any new variable (e.g. FULLSEND_SKILLS_DIR, fullsend-ai/agents#1672). That's a follow-up once the contract mechanism is agreed, so any new variable arrives with its contract surface already machine-checked.

    Adding FULLSEND_SKILLS_DIR can be done before this, so don't block that on this PR.