Agent Session Format Check

Agent Session Format Check

Answer one precise question for every supported agent:

Can current Agent Sessions code support the latest available session/storage/usage format from the latest available agent build?

The answer must be layered. Do not collapse version checks, stale samples, schema fingerprints, discovery contracts, and usage probes into a vague severity label.

Evidence-first:

• Gather a report + sample paths first.
• Do not change parsers/fixtures/docs without explicit user approval.

Related skill: agent-support-matrix — maintains the matrix YAML, ledger, and update-checklist workflow. This skill focuses on detection and evidence collection; agent-support-matrix focuses on recording and gating version bumps.

Process doc: docs/agent-support/monitoring.md — defines the compatibility verdicts, legacy severity model, cadence, and escalation workflow that feed into this skill.

1 Quick Start (all agents)

1. Run weekly monitoring:

   ./scripts/agent_watch.py --mode weekly

   Report path prints to stdout and is written under scripts/probe_scan_output/agent_watch/*/report.json.

2. In report.json, check each agent under results.<agent>:

   • compatibility.verdict, compatibility.scope, compatibility.blockers, and compatibility.next_action
   • verified_version, installed.parsed_version, upstream.parsed_version
   • compatibility.latest_status to distinguish current_fetch_known, cached_latest, and unknown latest-source states
   • weekly.local_schema (newest local session used for fingerprinting)
   • weekly.schema_diff and evidence.schema_matches_baseline
   • evidence.sample_freshness and evidence.fresh_evidence_source
   • compatibility.latest_real_session_failure when a prebump attempt failed
   • severity and recommendation only as legacy escalation fields

   Sanity-check version identity before acting on it. If a monotonic CLI reports an installed version newer than upstream.parsed_version, or the fetched tag belongs to a different release family (for example an SDK release instead of the CLI), treat the latest source as broken/unknown. Do not follow a generated prebump recommendation for an agent with no driver, and do not lower a verified ceiling to match a bad source.

3. Usage / limits reading (Codex + Claude) — always verify every weekly run. These drift independently of session schema (see §2), so a clean schema does not imply healthy usage reading. Each agent's results.<agent>.weekly.probes is a list; for every relevant entry confirm ok == true and exit_code == 0:

   • Codex — label == "codex_status_probe" (parse codex_status_json): the active CLI status channel (five_hour, weekly percent-left). The passive channel is the session JSONL token_count / rate_limits events, covered by the schema fingerprint above.
   • Claude — label == "claude_usage_probe" (parse claude_usage_json): the authenticated /usage reading (session_5h, week_all_models, week_opus). Also label == "claude_status" (parse claude_status_json): status.claude.com indicator/incidents. A failed or unparsed usage probe is a usage-format or auth regression even when versions match and the session schema is clean — never skip it, and report each probe's ok explicitly rather than collapsing it into the compatibility verdict.

4. Classify every new field for value, not just for safety (§1e). A field that parses cleanly is safe, which is not the same as handled. Do not close a drift finding until each new key or type has been called handled, noise, watch, or feature-candidate — and candidates filed in docs/backlog.md, not left as a remark in the ledger.

Interpretation:

• supports_latest: latest known build is covered by evidence.fresh_evidence_source == "latest_prebump_report" and compatibility.latest_real_session_evidence == true with compatibility.latest_status == "current_fetch_known". This verdict is not a clean bill by itself: also require the current weekly.schema_diff.unknown_types == [], weekly.schema_diff.unknown_keys == {}, healthy probes, and a passing discovery contract. The compatibility layer can reuse an older clean prebump report while a newer weekly sample has already drifted; in that case report and investigate the current drift rather than repeating supports_latest.
• supports_installed_only: installed build is covered by non-stale real local evidence, but latest is newer, cached from a prior report, unknown, or lacks fresh real-session proof.
• latest_unknown: no configured/reachable latest source or no real-session driver exists; do not claim latest support. Check STEWARDS.md before calling this a finding — for a steward-verified agent not installed here, this verdict is the expected output and the agent belongs to its steward, not to you (§1g).
• blocked_stale_sample: evidence predates the installed CLI; run prebump before claiming support.
• blocked_no_fresh_evidence: a version changed but no fresh matching sample proves support.
• format_drift_detected: unknown schema/storage/usage fields appeared; update fixtures/parsers.
• blocked_thin_sample: the sample was both narrow and tiny, so it evidenced nothing either way (§5). Generate a session that actually uses tools — not a one-line prompt.
• monitoring_broken: latest source, usage probe, or discovery contract failed.
• real_session_auth_failed in blockers: the real-session driver ran but the sandboxed agent was not authenticated; re-auth or provide the configured env token, then rerun prebump.

1a Real-Session Prebump Validation (required before latest claims)

Weekly scanning samples the newest on-disk session, which can predate a CLI upgrade and give a false "safe to bump" call (the codex 0.120.0 trap and the copilot session.shutdown trap). When weekly reports recommendation == run_prebump_validator — or before you stage any max_verified_version bump or latest-support claim — run the prebump path for every active agent being claimed. The driver exercises the currently installed CLI once inside a sandbox and diffs its output against the fixture baseline:

./scripts/agent_watch.py --mode prebump --agent codex --agent claude

Exit-code contract:

• 0 — every requested agent produced a fresh session and the schema matches baseline. Safe to bump.
• 2 — at least one fresh session's schema does not match baseline. Do not bump; investigate the schema diff in scripts/probe_scan_output/agent_watch/<slug>-prebump/report.json.
• 3 — at least one driver failed (timeout, auth, CLI not found, or discovery contract violation — wrong session root, wrong glob, or missing required event types).
• 4 — config/invariant error: unknown --agent (or one with no prebump block), missing/invalid discover_session contract, credential hygiene failure (oversize / mode), or sandbox breach (copilot hermeticity gate). Re-run with --allow-real-home only if you understand your real config dir will be mutated for that one invocation.

Keep the direct prebump report.json for every non-zero run and report its schema diff or driver error from that artifact. A later weekly scan may select older clean evidence or fall back to a stale local sample and omit the failed fresh diff; it does not erase what the just-completed driver observed.

Flags:

• --agent <name> (repeatable) — restrict to specific agents. An unknown agent or one without a prebump config block exits 4.
• --keep-sandbox — preserve the temp $HOME for debugging.
• --timeout-seconds N — per-driver timeout. CLI flag overrides per-agent config; falls back to config, then global default (120s).
• --force-fresh — suppress staleness evaluation for this run only (records stale_reason=forced_fresh in the report).
• --allow-real-home — copilot/real-HOME opt-in after a sandbox-breach diagnostic; never persistent.

A thin prebump can no longer downgrade a rich weekly union. A passing prebump replaces the weekly schema_diff with its own one-prompt session, and the thin-sample gate used to score only that: on 2026-08-13 codex reported blocked_thin_sample off a 20-event prebump while its weekly union carried 2808 clean events — adding evidence made the verdict worse. _sample_is_thin() is now applied to every available sample and only blocks when all of them are thin (pinned by test_thin_prebump_does_not_override_rich_weekly_union).

Give real_home_session agents a tool-using prompt. Their prebump session lands in the real store and enters the newest-_LOCAL_SCHEMA_SAMPLE_COUNT window, so a "Say hello" prompt actively degrades the next weekly sample — two such runs pushed antigravity's union down to 24 events and a genuine blocked_thin_sample. Claude's "Say hi, then use the Bash tool to run pwd." is the pattern; verify any prompt change with a real run, since a tool-using prompt can hang or return nothing on CLIs whose one-shot mode does not complete a tool turn.

A passing prebump is a floor, not a ceiling. Drivers use one-line prompts, so a fresh session may contain only the four most basic event types and still report fresh_matches_baseline=True — it proves the CLI still writes parseable output, not that rich event families are unchanged. For real_home_session: true agents that session also lands in the real store and becomes the newest sample; §5's multi-session union is what stops it from masking drift. Check the fresh session's type count before treating a pass as broad evidence.

Configured real-session drivers today are codex, claude, antigravity, copilot, opencode, hermes, openclaw, cursor, pi, kimi, and grok. Droid is legacy-only and excluded from active checks. Qwen, Devin, and fx have no drivers. Judge them on current weekly evidence plus their stewardship record (§1g), and never claim latest-build support without a session written by that build. Qwen cannot produce a newer session on this machine because its OAuth free tier was discontinued 2026-04-15; that remains blocked until a paid plan or alternate provider is configured (see §1c).

Staleness short-circuits the schema verdict. blocked_stale_sample is reported instead of drift, so a stale agent can be hiding real drift behind it. Kimi sat at blocked_stale_sample while every one of its sessions carried two unmodelled event types; building its driver surfaced them immediately. Treat a stale verdict as "unknown", never as "clean".

Prebump uses the hybrid env-var-first auth policy: if the relevant API-key env var (OPENAI_API_KEY, ANTHROPIC_API_KEY, FACTORY_API_KEY, GITHUB_TOKEN) is set it is forwarded into the sandbox and real HOME is never read. Otherwise the driver copies the declared credential file from real HOME into the sandbox after running three hygiene gates (64 KiB max, mode 0600, ≤90-day mtime warning). v1 drivers: codex_exec, claude_print, antigravity_print, copilot_prompt, opencode_run, hermes_oneshot, openclaw_local_agent, cursor_agent_print, pi_prompt, and kimi_prompt. Some OAuth/keychain-backed CLIs use real_home_session: true; run them with --allow-real-home so the session lands in the real agent store instead of copying single-use auth state into a sandbox.

1b Claiming a Newer Upstream Build (install → fresh session → verify → bump)

A version bump is never just a YAML edit. max_verified_version claims that the app parses sessions written by that build, and the only thing that can prove it is a session that build actually wrote. So when upstream.parsed_version is newer than installed.parsed_version, the sequence is fixed:

1. Install the newer CLI (its own updater, brew, or npm — whatever owns it).
2. ./scripts/agent_watch.py --mode prebump --agent <name>     # writes a FRESH session
3. Confirm exit 0 and fresh_matches_baseline == true.
4. Only now bump max_verified_version + ledger + tracking log (§6).

Step 2 is not optional, and skipping it makes the report actively worse. Installing an update rewrites nothing on disk: the newest session is still the one the old build wrote, so it is now older than the CLI binary and the agent flips from supports_installed_only (a real claim, backed by real evidence) to blocked_stale_sample (which per §1a means unknown, never clean). That is why the 2026-08-13 pass deliberately did not install Hermes 0.19.0: without a working one-shot run to generate a fresh session, updating would have destroyed a verdict it could not replace. Install only if you can complete step 2.

An agent with no prebump driver cannot do this loop at all. It is permanently capped at supports_installed_only, however clean its weekly looks, because nothing can produce a session from a build on demand. Build the driver first — grok_single (added 2026-08-17) is the worked example, and see §1d for what building one involves.

Re-read the installed version after the run. Several CLIs self-update because the prebump invoked them — copilot, cursor and antigravity all did on 2026-08-13 — so the build you validated may not be the build you started with. Bump to what the post-run report says is installed, not to what you intended to install.

1c When an Agent Cannot Produce a Session At All

Some agents are blocked for reasons no driver can fix, and the distinction matters because it decides whether there is engineering work to do:

• Fixable, ours — a broken install. OpenCode sat at installed=unknown with an Exec format error that read like a driver bug; /opt/homebrew/bin/opencode was actually the npm package's shell stub, because its postinstall never ran. Run file $(which <agent>) before suspecting anything schema-shaped.
• Fixable, theirs — a broken agent config. OpenClaw and Hermes both fail on borrowed backends that are misconfigured in the user's own environment (§ledger 2026-08-13).
• Not fixable by us — the account or plan is gone. Qwen's OAuth free tier was discontinued 2026-04-15; qwen -p returns "Run /auth to switch to Coding Plan, OpenRouter, Fireworks AI, or another provider," and qwen auth is itself "(removed)" in 0.21.x. No driver, re-login, or fixture can produce a session. Record the verdict honestly as blocked_stale_sample, say why in the matrix, and stop — it is a billing decision for the maintainer, not a task.

Do not let category three masquerade as category one. A missing driver looks like engineering work right up until you try to authenticate.

1d Building a Prebump Driver

Drivers live in scripts/agent_watch_prebump_drivers.py and register into DRIVERS; their config block is agents.<name>.prebump in docs/agent-support/agent-watch-config.json. Copy the nearest existing driver rather than starting from the Protocol.

Non-obvious requirements, each learned from a real failure:

• Declare discover_session with the modern roots/globs/required_types keys. The config gate rejects anything else with exit 4, and the runtime validator then proves the driver returned the artifact it claimed rather than some other file.
• Check the agent is in MATRIX_KEY_FOR_AGENT (scripts/agent_watch.py). A missing entry does not error: baseline_paths comes back empty, the diff takes its "no baseline → nothing diffs" branch, and the prebump reports fresh_matches_baseline=true having compared nothing. grok was missing from the prebump copy of that map on 2026-08-17 while present in the weekly copy, so its first driver would have passed vacuously. The map is now single-sourced and pinned by test_every_monitored_agent_is_registered_in_the_rebuild_tool.
• Prefer an API-key env var; fall back to a credential file. prepare_auth is the only auth path — never build env from os.environ inside a driver. Some agents have no key env var at all (grok's only documented one is GROK_SANDBOX), so the credential copy is their sole route and the 0600 hygiene gate always applies.
• Give it a tool-using prompt, and for real_home_session: true agents especially — their session lands in the real store and enters the newest-5 weekly window, so a "Say hello" prompt degrades next week's sample (§1a).
• If a session is a directory, verify the sidecars before returning ok. Grok's summary.json is a discovery precondition; returning a transcript whose sidecar never landed hands the fingerprinter half a session and reports the gap as schema drift.

Then verify the driver the way you would verify a finding: run it, confirm baseline_type_count in the report is non-zero (proving a baseline was consulted), and confirm the fresh session actually contains a tool call rather than four trivial events.

1e The Value Pass (what new fields are for, not just whether they break us)

This check has a structural blind spot, and it is worth stating plainly: it asks only "does it still parse?" Because the parsers read JSON as dictionaries, the answer is almost always yes — an unknown key is never read, an unknown type falls to a default: branch and becomes a .meta event. So every finding exits the funnel as "no parser change needed," which is true and also the end of the thought. Nothing ever asks whether upstream just started telling us something a user would want to see.

The cost is not hypothetical. Kimi's turn.ended.durationMs was noted in the 2026-08-13 ledger as "the natural source of a per-turn duration UI" and then sat there, because a ledger note is a remark, not work. And Qwen shipped with "usage/rate-limit tracking" listed as an unsupported surface while every one of its transcripts carried full per-call token accounting — 53 records in one ordinary session. The data was never missing. Nobody opened it.

So classify every new key or type into exactly one of four buckets, and record which:

Cheap heuristics for spotting a candidate: it is a number a user would ask about (tokens, cost, duration, context size), it names something currently anonymous (which connector ran this tool, which model, which provenance), or it records a state transition the UI hides (mode changes, rewinds, compaction).

Two rules that keep this honest:

• Prove handled, do not infer it from a similar field name. Cite the read path and a test whose behavior depends on the new shape. A parser's tolerant fall-through or raw JSON retention means safe, not handled.
• Check the value before believing the matrix. unsupported_surfaces describes what the app does, never what the agent emits. Qwen's entry made a present surface look absent for a full release. When a value pass contradicts a matrix line, the matrix line is the thing that is wrong.
• Measure before promoting. Claude's budget_usd carries real dollars and looked like a free replacement for the estimated-cost path — until counting showed it in 2 of 410 recent sessions, because it only appears when a budget is set. Confirm how often a field actually occurs, and on which record types, before filing it as a candidate.

Candidates go in docs/backlog.md under the matching area section, using that file's own entry format, with verified stamped to the date of the sweep that found them. A format check should end with two outputs: a clean bill of health, and a short list of things upstream started telling us that we are not yet using.

1f Steward Check (what a community steward runs)

An agent may have a steward: a contributor who uses that agent regularly and re-verifies its format a few times a year, or after a big vendor release. Maintained agents are checked by the maintainer; best-effort agents have no steward yet. A steward is not expected to know any of the above. They run one command:

./scripts/steward_check.py <agent>       # ./scripts/steward_check.py --list-agents

It runs the ordinary weekly scan (§1) restricted to that one agent, against the steward's own local sessions, and answers in plain sentences with one of three exits:

• 0 — all good. all good: <agent> format matches the baseline (N sessions sampled), plus the matrix's verified version. If their CLI is newer than the verified version and the schema still matches, it says the matrix entry can be bumped — that is the §1b evidence a maintainer needs.
• 1 — drift. The schema diff in plain words, a redacted sample written to scripts/probe_scan_output/steward_check/<agent>/redacted-sample/, and a ready-to-paste GitHub issue body (also saved as issue.md).
• 2 — cannot check. The CLI is not installed, there are no sessions on disk, or the repository has no baseline fixtures for that agent yet. Says which.

What it deliberately does not do: write or rebuild any baseline fixture. Deciding that drift is real and rebuilding a baseline stays a maintainer job (scripts/rebuild_stage0_baseline.py --agent <agent> --emit, §5/§7).

The sample reuses rebuild_stage0_baseline._redact — the same trimming that produces committed fixtures — and then re-scans the result for home directories, emails, key-shaped strings, IPs and long opaque ids. If anything survives, the sample is withheld entirely and the issue body says so; a steward is never handed an almost-clean file to paste into a public issue.

Maintainer side of the same command: when a steward's issue arrives, its sample is already in fixture shape, so it drops straight into Resources/Fixtures/stage0/agents/<agent>/ for a baseline rebuild. Verify the redaction yourself anyway before committing (§7).

1g Agents With a Steward (what the maintainer does not chase)

STEWARDS.md assigns each agent a tier: Maintained (the maintainer runs it daily), Steward-verified (a named contributor re-checks it on their own machine), or Best-effort (nobody yet). A sweep must read that tier before deciding an agent is a problem, because the scan cannot distinguish "this agent is broken" from "this agent is not installed here." Both surface as installed=unknown, latest_unknown, cli_binary_unresolved, or a discovery contract failing no_local_file.

For a steward-verified agent that is absent from this machine, all of those are the expected, correct output. They are not findings. Do not:

• report them as drift, gaps, or open work in the sweep summary;
• install the CLI to "close" them — a maintainer install with no sessions in it produces no better evidence and starts a support obligation nobody asked for;
• bump or lower max_verified_version from a machine with no sessions from that build. The steward's dated verification is the evidence, and overwriting it with a local guess destroys the only real record.

What a sweep does owe a steward-verified agent, in order:

1. State the tier next to the verdict. "fx: latest_unknown — steward-verified, no local CLI" is a complete, honest line. "fx: latest_unknown" alone reads as a hole.
2. Check the steward's date, not the schema. The question is not "did the format move?" — nobody here can answer that — it is "how long since anyone who runs this agent looked?" Two or three checks a year is the agreed cadence; past that, the entry is aging, and that is the finding.
3. Ping, don't investigate. When it is due, or when upstream shipped a release the record predates, ask the steward to run ./scripts/steward_check.py <agent>. That is a ten-minute job for them and an impossible one here.
4. Route what you already know. If the sweep saw something relevant — an upstream version the record predates, a contract that changed shape — put it in the ask so the steward is not re-deriving it.

Read the recommendation through the tier. The verdict vocabulary assumes the maintainer can run the agent, so for a steward-owned one it gives advice you cannot take. Translate: blocked_no_fresh_evidence / rec=run_prebump_validator on a steward agent does not mean "run prebump" — no driver here can produce a session for a CLI that is not installed. It means upstream moved past what the steward last verified, and the action is the ping in step 3. Configure the agent's upstream source even when nothing else about it is checkable locally: it is the only channel that can raise this, and without it the rule above is unenforceable. fx sat at upstream=unknown for exactly this reason while v0.0.6 and v0.0.7 shipped past a record reading 0.0.5, and the sweep reported nothing.

A steward-verified agent only becomes maintainer work when its steward goes quiet. At that point the honest move is to drop it to Best-effort in STEWARDS.md and say so, not to quietly keep claiming a verification nobody performed.

Ordering trap. STEWARDS.md is the human table and the matrix is the machine record; the file itself says the matrix wins on disagreement. But the tier lives only in STEWARDS.md — the matrix has no steward: field — so a purely matrix-driven sweep is structurally blind to stewardship and will keep re-reporting the same non-findings every week. Read both, and if you find yourself about to file a steward-verified agent as a gap, that is the bug.

2 Usage / Limits Drift (Codex + Claude)

Usage and limits tracking can drift independently of session schema. Monitor both.

Codex

• Passive channel: session JSONL token_count / rate_limits event structure. This is covered by the schema fingerprint only because codex is fingerprinted nested (§5) — these events live under event_msg.payload, and the flat fingerprint that ran until 2026-08-03 stopped at {payload,timestamp,type} and could never see them. If codex is ever moved back to the flat fingerprint, this channel goes unwatched again.
• Active channel (weekly): codex_status_capture.sh output schema — parsed as codex_status_json by agent_watch.py.
• Check: in results.codex.weekly.probes (a list), the entry with label == "codex_status_probe" returns ok == true and exit_code == 0. If not, investigate whether Codex changed its status output format.

Claude

• Active channel (weekly): claude_usage_capture.sh output schema — parsed as claude_usage_json.
• Context probe: ./scripts/claude-status --json records status.claude.com indicator/incidents (parsed as claude_status_json).
• Check: in results.claude.weekly.probes (a list), the entry with label == "claude_usage_probe" returns ok == true — if false, the usage API response format may have changed, or authentication may be required. The label == "claude_status" entry (parse claude_status_json) reports status.claude.com indicator/incidents.
• If probe health fails (parsing_failed, auth required, etc.), treat as high severity because the UI can break.

Known issue: Claude usage probe exit 16 (parsing_failed)

claude_usage_capture.sh can fail with exit 16 when Claude auth tokens are exhausted. The /usage TUI command itself stops working in this state. This is not a format change — resolution requires claude auth login. The failure is intermittent and clears after re-authentication. The Swift side silently retains the last known good snapshot.

What to look for in upstream release notes

• New or renamed fields in usage/billing/token responses.
• Auth changes (new scopes, cookie rotation, API key requirements).
• Rate-limit header changes or new quota enforcement mechanisms.

2a Model Price Freshness (Session Runway $ burn)

The runway's $ presentation prices per-type token rates against a model table, so it drifts whenever a provider changes prices or ships a model slug we don't know — with no schema change and no failing probe. Nothing else in this scan catches it. Unlike a broken probe, stale prices fail silently: the number still renders, just wrong.

Sources of truth (fetch these, don't recall them):

• Anthropic — https://platform.claude.com/docs/en/about-claude/pricing
• OpenAI — https://developers.openai.com/api/docs/pricing

The table lives in two places that MUST stay identical:

• docs/prices.json — served to clients from GitHub Pages (corrects shipped apps with no release)
• RunwayPriceTable.bundledJSON in AgentSessions/CodexStatus/RunwayPriceTable.swift — the compiled-in default (offline / pre-first-fetch)