fix(profile): preserve harness in profile diffs by drewstone · Pull Request #27 · tangle-network/agent-sdk

drewstone · 2026-07-09T22:03:21Z

What changed

add the existing AgentProfile.harness field to profile-diff set, remove, changed-axis, and prune behavior
centralize the property-axis list so these operations cannot drift independently
make pruning delete keys instead of writing undefined, preserving the base profile

Checked before building: AgentProfile.harness and the canonical AgentProfileDiff implementation already existed; this extends those surfaces rather than adding another profile type.

Verification

agent-interface: 40/40 tests
all 8 workspace package test, typecheck, and build commands
changeset status
git diff --check

tangletools

✅ Auto-approved drewstone PR — `7d47b36e`

This PR was opened by the trusted drewstone account.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: drewstone_author · 2026-07-09T22:04:41Z}

tangletools

🟢 Value Audit — sound


Verdict	sound
Concerns	0 (none)
Heuristic	0.0s
Duplication	0.0s
Interrogation	62.3s (2 bridge agents)
Total	62.3s

💰 Value — sound

Wires the previously-orphaned AgentProfile.harness field through the diff contract (set/remove/changed-axes/prune) and centralizes the property-axis list so the four sites can't drift again — a coherent, in-grain fix with no better approach available.

What it does: Adds harness as a first-class diff axis everywhere the contract enumerates axes: the removal type (profile-diff.ts:51) and its zod schema (profile-schema.ts:142), the applyRemoval deletion path (profile-diff.ts:213), and the axis-reporting/pruning loops. It replaces three separately-duplicated axis arrays (previously at profile-diff.ts ~257, ~278, ~318 in the #22 contract) with one sha
Goals it achieves: Closes a real drift gap: PR #22 added the diff contract without harness, PR #25 later added AgentProfile.harness, but the diff surfaces were never updated — so a diff could not target, report, or prune the harness axis and changedAgentProfileAxes would silently omit it. Centralizing the axis list prevents the next added profile field from drifting the same way. The delete-not-undefined chang
Assessment: Good on its merits and in the codebase's grain. It mirrors the explicit anti-drift philosophy already encoded in profile-schema.ts:214-228 (the MutuallyAssignable compile-time guard) by applying the same 'single source of truth' idea to the axis enumeration. Scope is proportionate — three files, behavior strictly additive plus the correctness fix for prune — and the new tests pin set-only, rem
Better / existing approach: none — this is the right approach. I searched for an existing axis list to reuse (grep for as const/axes/keyof AgentProfile across packages/agent-interface/src): the only as const arrays are unrelated (sandbox-size.ts:22, interaction.ts:170), and the three axis arrays this PR deletes were the only enumerations. A keyof AgentProfile derivation would not work: the diff deliberately
Model: opencode/kimi-for-coding/k2p7
Bridge attempts: 1

🎯 Usefulness — sound

Completes the existing AgentProfileDiff surface by adding the already-typed harness field to set/remove/changed-axis/prune, centralizing the axis list to prevent drift and fixing a base-clobbering bug in prune along the way.

Integration: Reachable as a public SDK contract: export * from "./profile-diff.js" at packages/agent-interface/src/index.ts:809 re-exports applyAgentProfileDiff, changedAgentProfileAxes, pruneAgentProfileDiff, defineAgentProfileDiff, and agentProfileDiffSchema. No internal repo caller consumes the diff surface yet (grep across all packages/agent-* turned up only the agent-interface package's own test file),
Fit with existing patterns: Fits the established pattern exactly. harness is treated identically to the other scalar-ish axes: set.harness flows through the existing mergeAgentProfiles overlay spread (agent-profile.ts:442-456, overlay wins); remove.harness = true mirrors remove.confidential = true and remove.identity = true (whole-axis clear, profile-diff.ts:213, 235); the schema entry `harness: z.literal(true).opt
Real-world viability: Holds up on the non-happy paths that matter. The delete set[axis] / delete remove[axis] change in pruneAgentProfileDiff (profile-diff.ts:296-310) is a genuine fix, not cosmetic: the previous set[axis] = undefined left the key present with value undefined, which the return guard Object.values(set).some(v => v !== undefined) tolerated but which would then be spread by mergeAgentProfiles as `
Model: opencode/zai-coding-plan/glm-5.2
Bridge attempts: 1

No concerns — sound change, no better or existing approach found. ✅

What this audit checks

It judges the change on its merits — not whether it was tasked out in an issue. Unticketed, fast-moving work is fine; the question is whether the change is good and whether a better or existing approach should be used instead.

Pass	What it asks
Heuristic	Vague title? Whitespace-only or cruft-bearing diff? (content signals only)
Duplication	Do added function/class names already exist elsewhere in the repo?
Value Audit	What does it do? What goal does it achieve? Is it good? Better architecture or already-exists?
Usefulness Audit	Does it integrate and fit? Will it hold up in real use and actually get used?

Findings are concerns, not blocks — the human reviewer decides what to do with them.

_{value-audit · 20260709T220608Z}

tangletools · 2026-07-09T22:10:42Z

✅ No Blockers — `7d47b36e`

Review health 100/100 · Reviewer score 82/100 · Confidence 70/100 · 5 findings (1 medium, 4 low)

	opencode-kimi	glm	deepseek	aggregate
Readiness	89	92	82	82
Confidence	70	70	70	70
Correctness	89	92	82	82
Security	89	92	82	82
Testing	89	92	82	82
Architecture	89	92	82	82

Reviewer score is advisory once the run is complete and the verdict has no blockers.

Full multi-shot audit completed 2/2 planned shots over 4 changed files. Global verifier still owns final merge decision. | Full multi-shot audit completed 2/2 planned shots over 4 changed files. Global verifier still owns final merge decision. | Full multi-shot audit completed 2/2 planned shots over 4 changed files. Global verifier still owns final merge decision.

🟠 MEDIUM Truthiness check on harness removal inconsistent with !==undefined guards — packages/agent-interface/src/profile-diff.ts

if (remove.harness) uses truthiness while all other new guards (model, permissions, tools, etc.) use !== undefined. For true | undefined this is functionally identical, but a maintenance hazard: if harness ever gained a falsy-but-defined value (e.g. false), this guard would silently skip. Use if (remove.harness !== undefined) for consistency with the pattern established on lines 195, 210, 214, 217, 220, 223, 226, 229, 232, 236, 239.

🟡 LOW No direct-apply test for set+remove of the same harness axis (remove-after-set) — packages/agent-interface/src/profile-diff.test.ts

The only test combining set.harness with remove.harness prunes before applying, so the documented remove-after-set semantics (applyAgentProfileDiff merges then removes, profile-diff.ts:260-266, interface comment :65-72) for the harness axis is never exercised directly. A diff like { set:{ harness:'codex' }, remove:{ harness:true } } applied to a profile with harness='claude-code' should deterministically yield harness=undefined (remove wins). Behavior is correct and documented, so non-blocking; add one assertion to lock it and guard the scalar axis the same way the record axes are implicitly covered.

🟡 LOW set+remove on the harness axis in one diff is not directly asserted — packages/agent-interface/src/profile-diff.test.ts

The remove-wins-over-base case is covered (line 170 'explicitly removes only the harness axis'), and the prune test (line 195) exercises a diff carrying both set.harness and remove.harness — but only AFTER pruning harness out. No test applies a diff with BOTH set.harness and remove.harness to confirm the documented 'remove deletes axes AFTER the overlay is applied' ordering for harness specifically. Tracing applyAgentProfileDiff→mergeAgentProfiles (overlay wins → merged.harness='codex') then applyRe

🟡 LOW AgentProfileDiffRemoval schema/interface lack a drift guard — packages/agent-interface/src/profile-schema.ts

profile-schema.ts:224-228 only enforces bidirectional assignability between agentProfileSchema and AgentProfile; there is no equivalent guard for agentProfileDiffRemovalSchema vs AgentProfileDiffRemoval (profile-diff.ts:46-63). This PR keeps them in sync for harness (literal(true)/true), so not a defect here, but a future axis added to only one side will compile clean and diverge silently. Pre-existing pattern, out of scope to fix in this shot; flagging for a follow-up.

🟡 LOW No MutuallyAssignable guard for AgentProfileDiffRemoval vs zod schema — packages/agent-interface/src/profile-schema.ts

The compile-time drift guard on lines 224-228 covers AgentProfile ↔ agentProfileSchema, but there is no equivalent guard for AgentProfileDiffRemoval ↔ agentProfileDiffRemovalSchema. Adding a new axis to AgentProfileDiffRemoval without the zod counterpart would not produce a compile error. This is pre-existing (all other removal fields have the same gap), and the centralized agentProfileDiffPropertyAxes array reduces the risk by making the axis list single-source, but a belt-and-suspenders type-level guard here would prevent future drift.

_{tangletools · 2026-07-09T22:10:39Z · trace}

tangletools

✅ Approved — 5 non-blocking findings — `7d47b36e`

Full multi-shot audit completed 2/2 planned shots over 4 changed files. Global verifier still owns final merge decision. | Full multi-shot audit completed 2/2 planned shots over 4 changed files. Global verifier still owns final merge decision. | Full multi-shot audit completed 2/2 planned shots over 4 changed files. Global verifier still owns final merge decision.

Full immutable report for this review: trace

Summary comment for this run: full summary

_{tangletools · 2026-07-09T22:10:39Z · immutable trace}

tangletools

✅ Auto-approved drewstone PR — `dba35912`

This PR was opened by the trusted drewstone account.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: drewstone_author · 2026-07-09T22:14:15Z}

tangletools

🟡 Value Audit — sound-with-nits


Verdict	sound-with-nits
Concerns	2 (2 weak-concern)
Heuristic	0.0s
Duplication	0.0s
Interrogation	195.5s (2 bridge agents)
Total	195.5s

💰 Value — sound-with-nits

Threads the new harness field through the profile-diff contract (set/remove/changed/prune) and centralizes the axis list so the three inline copies can't silently drop a field again — fixes two real latent bugs from PR #25's incomplete rollout.

What it does: Extends AgentProfileDiff so harness is a first-class diff axis: (1) adds harness to a new centralized agentProfileDiffPropertyAxes const that previously existed as three separate inline arrays in profile-diff.ts:268-310, none of which included harness; (2) adds harness?: true to AgentProfileDiffRemoval (profile-diff.ts:51) and the matching z.literal(true) to agentProfileDiffRemovalSchema
Goals it achieves: Make harness fully round-trippable through the diff/improvement-loop surface — consistent with PR #25's stated intent that harness is 'a first-class lever an improvement loop can optimize.' Concretely it fixes two latent gaps PR #25 left behind: changedAgentProfileAxes silently missed any harness change (the ablation/causal-analysis view was blind to it), and pruneAgentProfileDiff could not prune
Assessment: Coherent and in-grain. The scalar-removal handling clones the existing confidential pattern exactly (profile-diff.ts:213 vs :235), so no new shape is invented. The guards added around each applyRemoval axis (e.g. if (remove.model !== undefined)) are behavior-preserving — removeKeys/removeValues already short-circuit on undefined (profile-diff.ts:107, :120) — so they add clarity and avoid a red
Better / existing approach: Searched for an existing axis enumerator to reuse (git grep for keyof AgentProfile / profileAxes / ProfileAxis across all .ts) — none exists; only profile-diff.ts references these axes. Considered deriving the list from keyof AgentProfile to make drift structurally impossible, but it does not fit: AgentProfile intentionally mixes 'identity' axes (name/description/version/tags) with 'property' ax
Model: opencode/zai-coding-plan/glm-5.2
Bridge attempts: 2
Bridge warning: opencode/kimi-for-coding/k2p7: Bridge returned 503: {"error":{"message":"cli-bridge admission timed out after 30000ms","type":"admission_rejected","reason":"queue_timeout","admission":{"active":20,"queued":1,"maxActive":20,"maxQueue":48}}}

🎯 Usefulness — sound

Catches the profile-diff contract up to the recently-promoted AgentProfile.harness field (set/remove/changed/prune), centralizes the previously-duplicated 13-axis list to prevent recurrence, and fixes pruning to delete keys instead of leaving undefined-valued ghosts.

Integration: Fully reachable. The diff helpers are public exports (packages/agent-interface/src/index.ts:809 re-exports profile-diff.js); applyAgentProfileDiff/changedAgentProfileAxes/pruneAgentProfileDiff are the documented contract for 'a portable full-profile patch' (CHANGELOG 0.19.0) and harness is the named 'first-class lever an improvement loop can optimize' (CHANGELOG 0.20.0). This is a published SDK (@
Fit with existing patterns: Exactly in-grain. The codebase already chose an explicit property-axis list + a compile-time schema/interface drift guard (profile-schema.ts:219) as its pattern, and this PR honors it: the new agentProfileDiffPropertyAxes const (profile-diff.ts:8) replaces four copies of the same array and is the single place a future axis gets added. harness removal is true-only (scalar, like confidential), dis
Real-world viability: Holds up. Tests cover the load-bearing cases: set-only (164-167), remove-only with schema rejecting false (183-193), set-then-remove ordering (196-204), and pruning that isolates a co-set axis (206-231). The pruning switch from set[axis] = undefined to delete set[axis] is required for the .not.toHaveProperty('harness') assertion (line 226) and matters downstream for any consumer using `has
Model: opencode/zai-coding-plan/glm-5.2
Bridge attempts: 1

💰 Value Audit

🟡 Centralized axis list could be compile-time pinned to the interface [maintenance] ``

agentProfileDiffPropertyAxes (profile-diff.ts:8) is the single source of truth now, but nothing enforces it stays a subset of AgentProfile's keys. Adding as const satisfies readonly (keyof AgentProfile)[] would turn a future typo or a dropped field into a compile error — the same class of drift this PR is fixing. Pure hardening, not required to ship; the bidirectional MutuallyAssignable guard in profile-schema.ts:224 already catches schema/interface drift, just not diff-axis-list drift.

🎯 Usefulness Audit

🟡 Non-harness removal guards are behaviorally redundant [robustness] ``

profile-diff.ts:210-241 wraps every removeKeys/removeResources call in if (remove.X !== undefined), but removeKeys already early-returns on removal === undefined (line 107) and removeResources does the same (line 149). Only the harness guard (line 213) is load-bearing, because harness is a scalar handled outside removeKeys. The added guards on model/permissions/tools/mcp/subagents/resources/hooks/modes/metadata/extensions don't change any behavior; they obscure which single guard is actually

What this audit checks

It judges the change on its merits — not whether it was tasked out in an issue. Unticketed, fast-moving work is fine; the question is whether the change is good and whether a better or existing approach should be used instead.

Pass	What it asks
Heuristic	Vague title? Whitespace-only or cruft-bearing diff? (content signals only)
Duplication	Do added function/class names already exist elsewhere in the repo?
Value Audit	What does it do? What goal does it achieve? Is it good? Better architecture or already-exists?
Usefulness Audit	Does it integrate and fit? Will it hold up in real use and actually get used?

Findings are concerns, not blocks — the human reviewer decides what to do with them.

_{value-audit · 20260709T221911Z}

tangletools

✅ Auto-approved drewstone PR — `362b0bcb`

This PR was opened by the trusted drewstone account.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: drewstone_author · 2026-07-09T22:21:01Z}

tangletools

🟢 Value Audit — sound


Verdict	sound
Concerns	0 (none)
Heuristic	0.0s
Duplication	0.0s
Interrogation	103.7s (2 bridge agents)
Total	103.7s

💰 Value — sound

Closes a real gap: the harness preference (added to AgentProfile in PR #25) was unreachable from the diff contract and pruning silently destroyed base values; this wires harness through set/remove/changed-axes/prune and fixes prune to delete keys instead of writing undefined.

What it does: Adds harness as a first-class diff axis: harness?: true on AgentProfileDiffRemoval (profile-diff.ts:63) and agentProfileDiffRemovalSchema (profile-schema.ts:142); includes 'harness' in a new single source-of-truth axis list agentProfileDiffPropertyAxes (profile-diff.ts:11-26) consumed by the AgentProfileDiffAxis type (37-39), changedAgentProfileAxes (285/292), and pruneAgentProfileDiff (318). appl
Goals it achieves: (1) Let an improvement loop target harness routing as a lever — the explicit motivation in the AgentProfile.harness doc (agent-profile.ts:275-286) — which was impossible before because changedAgentProfileAxes never reported 'harness' and remove:{harness:true} failed both zod and the TS interface. (2) Make pruning preserve the base profile: previously pruneAgentProfileDiff(diff,['harness']) left se
Assessment: Good, in-grain, proportionate. harness is correctly classified as a property axis, not identity (consistent with its doc that it is a PREFERENCE, not IDENTITY). The centralization + exhaustiveness guard directly addresses the root cause (drift between the AgentProfile type and the diff contract) rather than patching the symptom. The delete-vs-undefined change is a genuine correctness fix, not cosm
Better / existing approach: none — this is the right approach. AgentProfileDiff and all its helpers live only in packages/agent-interface/src/profile-diff.ts (the canonical home; agent-profile.ts:7 notes sandbox merely re-exports), and grep confirms no other package reimplements diff/overlay/prune. The change extends that exact surface and reuses mergeAgentProfiles/applyRemoval rather than reinventing merge semantics. The on
Model: opencode/kimi-for-coding/k2p7
Bridge attempts: 1

🎯 Usefulness — sound

Adds the newly-introduced harness axis to the profile-diff system (set/remove/changed/prune) and centralizes the axis list with a compile-time exhaustive guard, fixing the gap left when AgentProfile.harness was added in PR #25.

Integration: The diff functions (applyAgentProfileDiff, changedAgentProfileAxes, pruneAgentProfileDiff) are public API exported at index.ts:809 and are the established mechanism for portable profile-improvement artifacts. Before this PR, the brand-new harness field (agent-profile.ts:287) was invisible to every one of them: changedAgentProfileAxes would not report it, pruneAgentProfileDiff would not strip it, a
Fit with existing patterns: Extends the one established pattern rather than competing with it. The centralized agentProfileDiffPropertyAxes array (profile-diff.ts:11) with the exhaustive compile-time guard (profile-diff.ts:32-35) is this codebase's own idiom — the same MutuallyAssignable/never-trick appears in harness.ts:78-82 and profile-schema.ts. Centralization removes the three-way duplication that let harness slip throu
Real-world viability: Holds up beyond the happy path. The most consequential change is switching prune from set[axis]=undefined to delete set[axis] (profile-diff.ts:320-321): an explicit undefined in the overlay would clobber the base profile's value when the pruned diff is re-applied via mergeAgentProfiles (shallow spread), whereas an absent key preserves it — the test at profile-diff.test.ts:229 asserts base harness
Model: opencode/zai-coding-plan/glm-5.2
Bridge attempts: 1

No concerns — sound change, no better or existing approach found. ✅

What this audit checks

It judges the change on its merits — not whether it was tasked out in an issue. Unticketed, fast-moving work is fine; the question is whether the change is good and whether a better or existing approach should be used instead.

Pass	What it asks
Heuristic	Vague title? Whitespace-only or cruft-bearing diff? (content signals only)
Duplication	Do added function/class names already exist elsewhere in the repo?
Value Audit	What does it do? What goal does it achieve? Is it good? Better architecture or already-exists?
Usefulness Audit	Does it integrate and fit? Will it hold up in real use and actually get used?

Findings are concerns, not blocks — the human reviewer decides what to do with them.

_{value-audit · 20260709T222629Z}

tangletools · 2026-07-09T22:30:47Z

✅ No Blockers — `362b0bcb`

Review health 100/100 · Reviewer score 86/100 · Confidence 70/100 · 4 findings (4 low)

	opencode-kimi	glm	deepseek	aggregate
Readiness	95	86	92	86
Confidence	70	70	70	70
Correctness	95	86	92	86
Security	95	86	92	86
Testing	95	86	92	86
Architecture	95	86	92	86

Reviewer score is advisory once the run is complete and the verdict has no blockers.

Full multi-shot audit completed 2/2 planned shots over 4 changed files. Global verifier still owns final merge decision. | Full multi-shot audit completed 2/2 planned shots over 4 changed files. Global verifier still owns final merge decision. | Full multi-shot audit completed 2/2 planned shots over 4 changed files. Global verifier still owns final merge decision.

🟡 LOW Patch bump may understate an additive feature — .changeset/tidy-runners-diff.md

The changeset bumps @tangle-network/agent-interface as 'patch', but the description ('Add the profile harness preference to AgentProfileDiff ... operations') describes adding a new preference to a diff set — typically an additive/minor change under semver. If the harness preference is a new public field on AgentProfileDiff, consider 'minor' to signal the additive capability to consumers. No runtime impact; judgment call for the maintainer.

🟡 LOW No test for harness set-overlay when base profile has no harness — packages/agent-interface/src/profile-diff.test.ts

baseProfile always sets harness: 'claude-code' (line 13). The 'sets only the harness axis' test (line 151) verifies overlay-wins (claude-code → codex), but no test covers the base-has-no-harness → set-harness path. This is the common real-world case (profiles created without a harness preference, then optimized to add one). The logic is correct (spread in mergeAgentProfiles handles absent fields), but the coverage gap means a regression in the merge path for harness wouldn't be caught. Add a test w

🟡 LOW Minor applyRemoval guard style inconsistency with confidential — packages/agent-interface/src/profile-diff.ts

harness uses if (remove.harness !== undefined) while confidential (line 247) uses if (remove.confidential). Both are functionally equivalent for true | undefined types, but inconsistent. Prefer the same guard style across both.

🟡 LOW Redundant undefined-guards in applyRemoval duplicate helper short-circuits — packages/agent-interface/src/profile-diff.ts

The new if (remove.model !== undefined) / if (remove.tools !== undefined) / etc. guards wrap calls to removeKeys/removeResources, which ALREADY short-circuit on removal === undefined (removeKeys line 119: if (!record || removal === undefined) return record). The guards are functionally no-ops for correctness — old code next.model = removeKeys(next.model, undefined) returned next.model unchanged; new code skips the assignment. The only measurable effect is avoiding one asMutable(next.tags) shallow copy when remove.tags is undefined ([line 208](https://github.com/tangle-network/agent-sdk/blob/362b0bcb977964f934798feafe1e95f35bc2f8ef/packag

_{tangletools · 2026-07-09T22:30:45Z · trace}

tangletools

✅ Approved — 4 non-blocking findings — `362b0bcb`

Full multi-shot audit completed 2/2 planned shots over 4 changed files. Global verifier still owns final merge decision. | Full multi-shot audit completed 2/2 planned shots over 4 changed files. Global verifier still owns final merge decision. | Full multi-shot audit completed 2/2 planned shots over 4 changed files. Global verifier still owns final merge decision.

Full immutable report for this review: trace

Summary comment for this run: full summary

_{tangletools · 2026-07-09T22:30:45Z · immutable trace}

fix(profile): preserve harness in profile diffs

7d47b36